Files
paper-how-to-use-this-techn…/research/04-sensing-alternatives.md
T

78 KiB
Raw Blame History

Sensing Alternatives for Pedestrian Footfall and Occupancy

Research note for the PulseCast paper — Heilbronn, November 2026

Scope. This note compares the main sensing approaches that could replace or complement BLE/Wi-Fi MAC sniffing for pedestrian footfall and occupancy measurement. It covers measured accuracy from peer-reviewed studies and official evaluations, planning-level cost ranges, privacy risk, deployment constraints, and the best use of each modality. It closes with a reference sensing stack and a calibration design for Heilbronn.

What this note does not do. It does not compare vendor products, does not treat vendor accuracy claims as evidence, and does not assume that a result from a controlled indoor test transfers to a Heilbronn street. Where a number comes from a single site or a small sample, that is stated. Costs marked (est.) are planning ranges derived from the cited hardware or labour anchors, not quotations.


Key findings

  1. BLE/Wi-Fi MAC sniffing is an index of discoverable radios, not a people counter. Only 5–12% of devices are discoverable via Bluetooth, Wi-Fi detection was 69% conditional on Wi-Fi being on in a vehicle-oriented field study, and MAC randomisation is defeatable for ~96% of Android phones (Kurkcu & Ozbay, 2017; Musa & Eriksson, 2012; Martin et al., 2017). It fails both the accuracy test and the privacy test.
  2. No single sensor dominates. Thermal arrays are the best default for narrow gateways (89–95% accuracy, low privacy risk, ~€150–400 per node). Edge cameras are the accuracy benchmark (MAPE 13.8% day / 29.8% night outdoors) but carry the highest privacy burden among the non-MAC options. mmWave and LiDAR fill wide or weather-exposed gaps; manual counts are the only direct ground truth.
  3. The "right way" is a tiered stack plus a Bayesian hierarchical calibration model. Fuse a small number of calibrated sensors, model sensor-specific detection probabilities, and report counts with credible intervals and calibration dates. Without calibration, the output is an index, not a measurement.
  4. For Heilbronn: 8–12 thermal nodes at narrow gateways, 2–4 edge cameras at wide plazas, 1–2 mmWave/LiDAR gap-fillers, twice-yearly manual/video ground truth, and transit validators as covariates. No MAC sniffing, no face recognition. A 12-week, six-site pilot with explicit go/no-go thresholds (MAPE ≤ 20%, |bias| ≤ 10%, interval coverage ≥ 75%).
  5. Cost and legal reality: a year-1 pilot is roughly €81k–161k (estimate), with €30k–50k recurring; calibration is the largest recurring cost. A DPIA under Art. 35(3)(c) GDPR, BDSG §4 compliance, and a no-raw-video/no-raw-point-cloud design are preconditions, not downstream tasks.

1. Comparison table

How to read the accuracy column. "Accuracy" is not one metric. A study may report exact-count accuracy, mean absolute percentage error (MAPE), mean absolute error (MAE), F1 for presence detection, or correlation with a reference. These are not interchangeable. A method that is excellent at "is anyone here?" can be poor at "how many people are here?". The context in the cell matters more than the headline number.

Modality What it measures Reported performance (source) Cost (2026 planning estimate) Privacy risk Main deployment constraints Best use
BLE MAC sniffing (baseline) Discoverable Bluetooth radios, not people 5–12% of devices discoverable via Bluetooth; 97%/92% count accuracy only when penetration is assumed 100% (Kurkcu & Ozbay, 2017). MAC randomisation defeated for ~96% of Android phones (Martin et al., 2017). €100–500 Very high — MAC addresses are personal data; persistent tracking; re-identification risk Unknown and changing device penetration; multiple devices per person; staff and passers-by indistinguishable; packet loss; legal basis difficult None for public footfall. At most a short-lived, calibrated index under a DPIA.
1. Edge AI camera People crossing a line or occupying an area Outdoor 18-site evaluation: MAPE 13.8% day, 29.8% night; best-decile 5.3%/7.3%, worst-decile 30.5%/62.8% (Cyganek et al., 2022). Overhead test: 97% live, 92% over two days (Ahamed et al., 2024). Official review: 2–14% error for video imaging (Ozan et al., 2021). €500–2,000; commercial systems €1,100–€7,400 High — identifiable images even without face recognition; DPIA required Lighting, occlusion, rain, mounting height, field of view, compute/power, public acceptance Reference and validation at 2–4 complex sites; wide plazas; pedestrian/cyclist classification.
2. Thermal / IR array Warm bodies crossing a doorway; direction 93% average occupancy accuracy (Mohammadmoradi et al., 2017); 89–92% doorway accuracy (Maaspuro, 2020); 95% at normal walking speed up to 30 °C, 100% in stable 22–26 °C (Perra et al., 2021). Degrades above 30 °C, beyond ~7.5 m, with fast walking or groups. €150–400 (sensor ~$50) Low — low-resolution heat maps, not photographic images; aggregate at edge Narrow doorways only; ambient temperature; mounting height; group counting; limited range Default permanent counter at store entrances, narrow passages, transit shelters.
3. mmWave radar Moving people as point clouds; presence, flow, count Outdoor Amsterdam: 96% presence, 69% count, −2.5% aggregate bias (Vaidya & Zúñiga, 2024). Three-month outdoor: F1 84%, flow correlation 60–85% (Wang et al., 2023). Single radar: MAE 1.53 up to 21 people (Pallaprolu et al., 2024). Festival: correlation 82.9–94.2%, ±50% small crowds / ±25% large crowds (Sakhnini et al., 2024). €300–700 ($311 node; $120.87 sensor) Low — point clouds, no images; raw data can still track individuals Rain, foliage, multipath, groups, range ~8 m, mounting, power Wide, weather-exposed crossings; transit stops; presence and flow, not exact occupancy.
4. LiDAR 3D geometry of people; tracking and direction 2D LiDAR: >97% disaggregate count; 1.1% false direction; 0.7% overcount; 1.3–2.7% undercount; 1.6–4.3% AAPD at 15 min (Lesani et al., 2020). Official review: 5% or more average error in crowded scenes; $8,000 unit (Ozan et al., 2021). €1,000–8,000; low-cost 2D from ~$100 Low–moderate — geometric point cloud; individual tracking possible Cost; eye safety; weather; mounting; calibration; power High-accuracy fixed-site counting; 1–2 reference sites, not a city-wide network.
5. Wi-Fi CSI Channel perturbations caused by bodies; device-free 80% accuracy up to 7–8 people (Di Domenico et al., cited in Brena et al., 2021); 80% perfect-or-±1 up to 30 people (Xi et al., cited in Brena et al., 2021); 99% with six links in controlled rooms (Brena et al., 2021); 92.7% with RTT in one 80 m² office (Yonekura et al., 2025); 61% cross-domain (VICount, cited in Yonekura et al., 2025). €200–1,000 per link; €10k–30k engineering Moderate–high — device-free but can infer presence, movement, occupancy; GDPR still applies Per-room calibration; furniture and crowd changes; CSI extraction hardware (Intel 5300 no longer current); no standard API; poor cross-domain transfer Indoor rooms where cameras are unacceptable and thermal cannot cover; research pilots only.
6. Wi-Fi probe sniffing Probe requests and MAC addresses 69% detection of passing phones if Wi-Fi is on; 67 m mean trajectory error; ~1 device per 5 vehicles (Musa & Eriksson, 2012). No stable people multiplier. Randomisation defeated for ~96% of Android phones (Martin et al., 2017). €100–500 Very high — MAC addresses are personal data; tracking; DPIA; ePrivacy Wi-Fi-on rate; randomisation; probe timing; multiple devices; legal basis; public trust None for public footfall. The privacy cost is not justified by the data quality.
7. Telco network data Serving-cell events / CDR; aggregated mobility Population mapping comparable to census-based methods (Deville et al., 2014). Spatial resolution: 250–1,000 ft urban, 1,500–3,000 ft suburban, 1–3 miles rural; not suitable for intersection-level analysis (BTS, 2025). Urban subscribers typically within ~500 m of a cell tower (Flowminder, n.d.). No node; bespoke licence, €10k–100k+/year (est.) High — individual mobility traces; aggregation and re-identification risk Operator cooperation; market share; cell-sector resolution; latency; licensing; regulatory approval City-wide context, origin–destination, tourism and event flows — not street-level footfall.
8. GPS probe / map providers App-based GPS traces; route and volume estimates Strava users biased toward cyclists (+8%), males (+15.7%), ages 35–54 (+20.4%) (Venter et al., 2023). Model accuracy ±32–160 cyclists on 75% of streets (Strava Metro, n.d.). Pedestrian data are rare (Lee & Sener, 2020). No node; Strava ~$0.55/user/year (2021); commercial feed €10k–50k/year (est.) High — individual traces; commercial re-use; GDPR Sample bias; indoor gaps; urban canyon; app penetration; licensing Cyclist route and origin–destination analysis; recreational trends. Not inner-city footfall.
9. Transit smart cards / APC Fare validations, boardings, turnstiles Transaction-level for validations, but only for passengers who validate. Fare evasion: 5% partial / 37% bus-only in one system (Munizaga et al., 2020); 20–50% in some California agencies (Caltrans, 2025). €0–5k/year data extract if AFC exists; €5k–15k per new validator/gate (est.) High — individual trip records; pseudonymous but linkable Only transit users; fare policy; operator cooperation; no pedestrian-zone coverage Transit demand covariate; commuter and event signal. Not general footfall.
10. Manual counts / video annotation Direct human observation; ground truth Manual clicker average error 11% (8% and 15% on two dates) (Diogenes et al., 2007). Observers accurate for about 2 hours; 2-hour-based annual estimates can be off by up to 60%; one week of counts recommended (Nordback, 2014). €25–40 per observer-hour loaded (est.); campaign €11k–18k Low for counts; high for video annotation (faces) Labour; training; weather; sampling design; inter-rater reliability Ground truth and calibration; regular audits; not continuous monitoring.
11. Sensor fusion / Bayesian hierarchical model Latent true count, combining sparse modalities No single accuracy figure. Hierarchical models combine multiple estimates and propagate uncertainty (Parsons et al., 2022); Bayesian spatio-temporal models map pedestrian traffic from sparse counts (Zaouche & Bode, 2023). Performance depends on calibration data and model checks. €20k–60k development; €5k–15k/year maintenance (est.) Depends on inputs — aggregated outputs can be low-risk; raw inputs inherit their risks Data engineering; identifiability; drift; validation; explainability The core of a defensible stack: fuse thermal, camera, radar, manual and transit data.

Cost basis. Cost ranges are 2026 planning estimates in euros, excluding VAT, installation, backend and maintenance unless stated. They are anchored to published figures: a Grid-EYE sensor at ~$50 (Perra et al., 2021), with a complete thermal node estimated at ~€150–400; a 60 GHz radar node at $311, itemised as $120.87 sensor + $49.77 Raspberry Pi 4B + $101.12 PTFE radome + $12.74 back casing + ~$26.50 misc (Wang et al., 2023); commercial radar nodes at ~$600 (Wang et al., 2023); video systems at $1,200–$8,000, depth cameras at $9,900–$12,330, thermal cameras at $4,800 and laser scanning at $8,000 (Ozan et al., 2021); low-cost 2D LiDAR from ~$100 (Umezu et al., 2023). Where a source is in USD, the euro range is a rounded planning conversion, not an invoice.

Evidence quality. The strongest evidence in this table is the outdoor camera evaluation (18 sites, 207 minutes, manual reference), the 2D LiDAR evaluation (two sites, disaggregate and 15-minute counts), the three-month outdoor radar deployment, and the official NCDOT technology review. The weakest are the single-room CSI results and the short, controlled radar tests; the three-month radar deployment is stronger but still reports wide outdoor error. No study in this table is a randomised multi-city trial of the kind that would justify a single national accuracy figure.


2. Modality notes

2.0 The baseline: why BLE MAC sniffing is not a people counter

Principle. A BLE scanner listens for advertising packets. Each packet carries a device address; the scanner counts unique addresses in a time window. The same logic applies to Wi-Fi probe requests. The output is an index of discoverable radios, not a count of people.

Measured accuracy. The fundamental problem is penetration, not classifier quality. Kurkcu and Ozbay (2017) report that 5–12% of devices are discoverable via Bluetooth. Their own pedestrian-counting tests reached 97% and 92% accuracy only when the penetration rate was assumed to be 100%; they explicitly warn that the method depends on the ratio of people with discoverable devices. A separate evaluation cited by Kurkcu and Ozbay (2017) found Bluetooth detection as low as 2.0%; a combined Wi-Fi/Bluetooth system had 15% mean and 9% median pedestrian-flow prediction errors. There is no stable multiplier from devices to people: penetration varies by device brand, OS, battery state, privacy settings, age, and location.

Randomisation. Modern phones randomise MAC addresses, but the randomisation is imperfect. Martin et al. (2017) defeated it for approximately 96% of Android phones using probe-request fingerprints, and found a low-level control-frame flaw that applied to 100% of tested devices. In other words, the "privacy-preserving" randomisation is not a reliable anonymisation layer, and the "count" is not a reliable people count. The two failures point in opposite directions: the data are both biased and identifiable.

Cost. Very low: a single-board computer or consumer access point can capture packets (Musa & Eriksson, 2012, used $70 access points). The low hardware cost is why the method spread. It is not the relevant cost: legal review, DPIA, signage, retention controls and public consultation dominate.

Privacy. MAC addresses are personal data. The Dutch Data Protection Authority states this explicitly and treats Wi-Fi and Bluetooth tracking as processing of indirectly identifying personal data; municipalities may use it only under strict conditions, for specific periods and precisely described areas (Autoriteit Persoonsgegevens, 2025). In Germany, the same conclusion follows from Art. 4(1) GDPR and the official guidance of the Schleswig-Holstein data protection authority (Unabhängiges Landeszentrum für Datenschutz Schleswig-Holstein, 2017). Hashing does not solve the problem if the hash is stable and linkable over time. The GDPR's purpose limitation, data minimisation, storage limitation and DPIA obligations all apply; a large-scale systematic monitoring of a publicly accessible area triggers Art. 35(3)(c) (GDPR, 2016; EDPB, 2020).

Deployment. The method requires no cooperation from the public, which is exactly why it is attractive and exactly why it is contested. It also requires a legal basis that most city marketing or retail analytics purposes do not have. The project's own data notes make the same point: the PAX series is an index of estimated devices, the conversion to people is undocumented, and device penetration is unknown (PulseCast data notes, 2026).

Verdict. BLE/Wi-Fi MAC sniffing is not a measurement technology for public footfall. It is a low-cost radio-presence index with unknown and unstable coverage and a high privacy cost. The paper's critique is correct; the rest of this note is about what to use instead.


2.1 Camera-based people counting (edge AI, no face recognition)

Principle. A camera frames a doorway, corridor or plaza. A person detector (YOLO-family, SSD-MobileNet, or a newer transformer) locates people; a tracker associates detections across frames; a line-crossing or region-transition rule counts entries and exits. Edge inference means the video never leaves the device: the camera outputs counts, directions and confidence, not frames.

Measured accuracy. The most relevant outdoor evaluation is Cyganek et al. (2022): 18 locations, 207 minutes of real footage, almost 1,000 pedestrians, manual annotation as reference. Mean MAPE was 13.8% in daylight and 29.8% at night. The distribution matters more than the mean: the 9th-decile error was 5.3% by day and 7.3% by night, while the 1st-decile error was 30.5% and 62.8%. The reference is manual annotation, and the authors note that the manual counting error is unknown, so the MAPE is a comparison against an imperfect ground truth. In other words, a well-placed camera in good light can be within a few percent; a badly placed camera in the dark can miss a third or more of the crowd. Ahamed et al. (2024) report 97% accuracy in live overhead tests and 92% over a two-day continuous test, with 20–27 FPS on an Intel NUC 12 Pro and no GPU. That is a strong result for a controlled doorway, not a general outdoor guarantee. The NCDOT review reports 2–14% error for video imaging systems, with the caveat that complex intersection movements reduce accuracy (Ozan et al., 2021).

Cost. The NCDOT review puts commercial video-imaging systems at $1,200–$8,000 per unit and depth cameras at $9,900–$12,330 (Ozan et al., 2021). A do-it-yourself edge node — industrial camera, Jetson-class compute, weatherproof housing, PoE — is roughly €500–2,000 in hardware (est.). Installation, mounting, power and network add €200–500 per site (est.). The hidden cost is model validation: a generic detector trained on COCO or a European street scene will not be equally accurate on a rainy Heilbronn Saturday; local annotation and threshold tuning are part of the project.

Privacy. This is the highest-privacy-cost option among the non-MAC alternatives, even without face recognition. Frames containing identifiable people are personal data under Art. 4(1) GDPR. The EDPB's video-device guidelines explicitly discuss "simple counting algorithms" as less intrusive than biometric analytics, but they do not exempt them from the GDPR: lawfulness, necessity, proportionality, transparency and storage limitation still apply, and a DPIA is required for systematic monitoring of a publicly accessible area on a large scale (EDPB, 2020; GDPR, 2016, Art. 35(3)(c)). In Germany, §4 BDSG adds a necessity/proportionality test for video surveillance of publicly accessible spaces and requires deletion when the purpose is achieved (BDSG, 2017). The AI Act does not ban people counting as such; it prohibits real-time remote biometric identification in publicly accessible spaces for law enforcement, with limited exceptions (AI Act, 2024, Art. 5(1)(h)). Adding face recognition or re-identification would change the legal analysis and should be treated as out of scope.

The practical mitigation is architectural: inference on the device, frames discarded immediately, only counts and confidence transmitted, no remote access to the video stream, signed firmware, encrypted storage for the short buffer needed by the tracker, and a published retention policy (ideally zero frame retention). Even then, a camera in a public square is a visible intrusion; signage and a public explanation are not optional.

Deployment. Cameras need a clear line of sight, stable mounting, power and network. They struggle with low light, strong backlight, rain and snow, dense occlusion, and people carrying umbrellas or large bags. They are also the modality most likely to attract public opposition. In a city centre, the acceptable use is sparse: a few reference sites where the accuracy benefit justifies the privacy cost, not a blanket network.

Best use. Reference and validation at 2–4 complex sites: wide plazas, intersections, event spaces, and locations where thermal arrays cannot cover the width. Cameras are also the best tool for classifying pedestrians versus cyclists and for validating the thermal and radar counts.

Verdict for Heilbronn. Use a small number of edge cameras as the accuracy benchmark and for wide, high-value locations. Do not make them the default. The privacy and governance burden is real, and the accuracy advantage over a well-placed thermal array is modest at a doorway.


2.2 Thermal / IR array sensors

Principle. A thermopile array (typically 8×8, 32×24 or 80×60 pixels) measures infrared radiation and produces a low-resolution heat image. A person crossing the field of view appears as a moving warm blob. On-device processing thresholds the blob, tracks its direction, and increments an in/out counter. No visible image is formed; at 8×8, a person is a few pixels, not a recognisable face.

Measured accuracy. The doorway literature is consistent. Mohammadmoradi et al. (2017) report 93% average accuracy for room occupancy, with 96%/94% at a 180 cm door and 100%/96% at a 90 cm door; in uncontrolled classrooms and labs, accuracy was 89–92%. Maaspuro (2020) reports 89–92% for a doorway occupancy counter. Perra et al. (2021) report 95% average detection at normal walking speed for ambient temperatures up to 30 °C, 90% or better for brisk walking up to 26 °C, and 100% in a one-month stable 22–26 °C test. The failure modes are well characterised: accuracy falls with distance (1.5 m, 5 m, 7.5 m), with walking speed (normal, brisk, running), and when ambient temperature approaches body temperature (above 30 °C). At 7.5 m and 30–34 °C, detection can fall to zero. A single 8×8 sensor covers roughly 2.5 m × 2.5 m at the best mounting height; wider doors or plazas need multiple sensors or a higher-resolution array.

Cost. The Grid-EYE sensor is about $50 (Perra et al., 2021). A complete node — sensor, microcontroller, radio (LoRaWAN, Zigbee, Wi-Fi), power, housing — is roughly €150–400 (est.). A high-end thermal camera is a different product: the NCDOT review lists a FLIR camera and board at $4,800 (Ozan et al., 2021), which is unnecessary for counting at a doorway.

Privacy. This is the lowest-risk optical modality. At 8×8 or 32×24, the data are low-resolution heat maps, not photographic images; they are much harder to link to an individual. The EDPB's video guidelines are about devices that produce images of identifiable people; a low-resolution thermopile array that outputs counts is much easier to justify. That does not mean "no GDPR": if raw frames are stored and linked to a person, they can become personal data. The safe design is to process on the device, output only counts and direction, and never store raw frames. A DPIA is still prudent, but the risk assessment is simpler and the public explanation is easier ("a heat sensor counts entries; it cannot see you").

Deployment. Thermal arrays are best at narrow, controlled passages: store entrances, doorway thresholds, corridor mouths, shelter entrances. They need a mounting height of roughly 2.3–3.5 m, a clear view of the floor, and protection from direct sunlight and heat sources (radiators, heaters, ovens). Ambient temperature is the key constraint: in a heatwave, a 30 °C+ street and a 37 °C body are hard to separate. Groups are another constraint: two people walking shoulder-to-shoulder can merge into one blob. Directional counting is reliable for single-file passage; it degrades in dense, bidirectional crowds.

Best use. The default permanent counter for Heilbronn's narrow gateways: store entrances, the passages around Stadtgalerie, Marahaus and Götzenturm, and transit shelter entrances. Pair with a second sensor or a camera where the passage is wider than ~2.5 m.

Verdict for Heilbronn. The best privacy/cost/accuracy trade-off for the inner-city's narrow entries. It is not a plaza sensor. Use it where the geometry is on your side.


2.3 mmWave radar and LiDAR

Principle. A 60 GHz frequency-modulated continuous-wave (FMCW) radar transmits chirps and measures range, velocity and angle from the reflections. People appear as moving point clouds; clustering and tracking turn the cloud into tracks. LiDAR does the same with laser pulses, producing a geometric point cloud with centimetre-to-decimetre angular resolution. Neither modality forms a photographic image.

Measured accuracy. The outdoor evidence is more nuanced than the indoor literature. Vaidya and Zúñiga (2024), working with the municipality of Amsterdam, deployed a radar across a street and compared it with a camera: 96% accuracy for presence detection, 69% for people counting, and an aggregate undercount of only 2.48% (1,005 true vs 980 estimated) over the evaluation. The reference is a camera, so the comparison inherits the camera's own error. Wang et al. (2023) ran a three-month outdoor deployment and report 85% accuracy for detecting empty spaces, an F1 of 84%, and flow correlations of 60–85% depending on weather. They conclude that outdoor mmWave still needs improvement compared with cameras (error below 10% in their framing). Pallaprolu et al. (2024) report an MAE of 1.53 for crowds of up to 21 people using a single radar and a shadowing model, over 22 experiments in four areas. Sakhnini et al. (2024) evaluated a festival crowd of 0–1,500 people and found 82.9–94.2% correlation with bracelet-based access counts, with 95% prediction intervals of ±50% for small crowds and ±25% for large crowds. Ren et al. (2023) report 96.25% average tracking precision for grouped people in outdoor scenarios, but that is a tracking metric, not an exact crowd count. Indoor results are stronger, but they do not transfer unchanged to a street.

LiDAR has the strongest fixed-site counting evidence. Lesani et al. (2020) evaluated a 2D LiDAR system at two high-volume locations: more than 97% of pedestrians counted at the disaggregate level, 1.1% false direction, 0.7% overcount, 1.3% and 2.7% undercount, and 1.6%/4.3% average absolute percentage deviation at 15-minute aggregation. That is better than most camera results, but it comes from a carefully installed system at two sites. The NCDOT review reports an average error of 5% or more for laser scanning in crowded scenes and an $8,000 unit price (Ozan et al., 2021).

Cost. The 60 GHz radar node is the cost standout: Wang et al. (2023) itemise $311 for sensor, Raspberry Pi, radome, casing and misc; a commercial indoor node was $600. LiDAR is more expensive: low-cost 2D modules start around $100 (Umezu et al., 2023), but a deployment-grade outdoor unit with weatherproofing, mounting and processing is €1,000–8,000 (est.), and 3D LiDAR is higher. Installation and calibration add €300–800 per site (est.).

Privacy. Both modalities are privacy-preserving relative to cameras: they produce geometry, not photographs, and the raw signal does not contain facial features. That does not make them anonymous in all cases. A point cloud can track an individual through a space, and a persistent track can be linked to other data. The GDPR still applies if the data can be related to a person; the safe pattern is the same as thermal — process at the edge, output counts and tracks only for the minimum time, do not store raw point clouds, and document the purpose. mmWave can also pick up small movements, including breathing in some configurations, which is useful for stationary occupancy but raises the possibility of inferring more than presence; keep the inference scope narrow.

Deployment. mmWave at 60 GHz is licence-free in the ISM band and has a range of roughly 8 m in the outdoor deployments cited here. It is sensitive to rain, foliage, multipath and moving clutter; a wet street or a swaying tree can produce false points. It undercounts groups because people close together merge into one cluster. LiDAR is more accurate but more expensive, has moving parts in many units, and needs a clear, stable mounting with an unobstructed view. Both need power and backhaul.

Best use. mmWave is a good gap-filler at wide, weather-exposed crossings and transit stops where a thermal array cannot cover the width and a camera is politically difficult. Use it for presence and flow, not for exact staffing counts. LiDAR is best at one or two high-value reference sites where its accuracy justifies the cost.

Verdict for Heilbronn. Use one or two mmWave nodes at wide, exposed locations (Hagenbuchersee, Neckarufer) and, if the budget allows, one LiDAR reference site. Do not build the network on radar; the outdoor count error is too high for staffing without calibration.


2.4 Wi-Fi CSI sensing / device-free sensing

Principle. Wi-Fi channel state information (CSI) describes how the wireless channel distorts a known signal across subcarriers. People moving in the channel change the amplitude and phase; a classifier maps those changes to a crowd count. The method is "device-free" in the sense that the people do not need to carry a device, but the sensing hardware is a Wi-Fi transmitter and receiver.

Measured accuracy. The early results are impressive in controlled rooms. Brena et al. (2021) summarise the state of the art: Di Domenico et al. report about 80% accuracy for up to 7–8 people; Xi et al. report an 80% probability of a perfect count or a one-person error for up to 30 people. Their own multi-link approach reaches 99% with six links in three rooms, versus 44% with one link. Zhou et al. (2020) report a counting error within one person for 100% of cases at a crowd size of 6 and 97.7% at a crowd size of 34. More recent work reports 92.7% with round-trip-time measurements in a single 80 m² office (Yonekura et al., 2025) and 85% for 0–10 people in VICount (cited in Yonekura et al., 2025). But the same literature documents the weakness: environment dependence. Santos et al. (2024) collected data over three months in multiple rooms and found that a model trained without knowledge of the target environment performed substantially worse; partial training on the target environment improved classification accuracy by about 16–20%. Yonekura et al. (2025) report 61% cross-domain accuracy for VICount and note that CSI lacks standardisation, which limits practical deployment.

Cost. The radio hardware is cheap: a transmitter, a receiver and two antennas. The expensive part is engineering. CSI extraction is not exposed by all commodity access points; the widely used Intel 5300 NIC is no longer current; different chipsets expose different formats; and each room needs calibration. A realistic pilot budget is €10k–30k for engineering, plus €200–1,000 per link (est.). This is not a plug-and-play sensor.

Privacy. CSI does not require MAC addresses, which is a genuine advantage over probe sniffing. But it is not automatically anonymous: the signal changes with presence, movement, and even breathing, and a trained model can infer occupancy patterns over time. If the data are linked to a household or a specific room, they can be personal data. The GDPR's purpose limitation and minimisation principles still apply; a DPIA is appropriate for a building-scale deployment. The privacy story is better than MAC sniffing, but it is not "privacy-free".

Deployment. CSI is best indoors, in rooms or corridors with stable geometry and a known transmitter/receiver layout. It needs a calibration dataset per environment, and it drifts when furniture, doors, or the crowd's behaviour change. It is poor at distinguishing people from other moving objects (carts, doors, pets) and at counting dense crowds. Outdoor use is rare and difficult because of multipath, weather and uncontrolled geometry.

Best use. Indoor occupancy in offices, libraries, care facilities or transit waiting rooms where cameras are unacceptable and a thermal array cannot cover the room. It is a research-grade modality, not a city-scale footfall sensor.

Verdict for Heilbronn. Not ready for the inner-city footfall network. Keep it as an indoor research option if a building-level occupancy use case appears.


2.5 Wi-Fi probe request sniffing

Principle. Wi-Fi devices periodically send probe requests to discover access points. Each request carries a MAC address and information elements. A passive monitor records the requests, deduplicates by MAC, and counts unique devices in a time window.

Measured accuracy. Musa and Eriksson (2012) is the canonical field study. In a 12-hour trial on a road with 37,000 average daily traffic, a single roadside monitor detected 7,000 unique devices, or about one device for every five vehicles. Across their deployments, a monitor detected a passing smartphone 69% of the time if Wi-Fi was turned on, and their trajectory estimates had a mean error of 67 m against GPS. The conditional is doing a lot of work: Wi-Fi-on rates vary, and the method cannot see devices with Wi-Fi off, devices in power-saving mode, or people without phones. Those deployments were vehicle-oriented, so 69% is not a pedestrian detection rate; it illustrates the conditional nature of the measurement. The count is an index of emitting devices, not people. The same study calls the method "relatively benign" because it only tracks phones in an area of interest, and suggests seeded hashes as a further protection; that is a research-era judgment, not a modern legal position.

Randomisation. Martin et al. (2017) show that MAC randomisation is widely deployed but poorly implemented: devices sometimes send their true global address when they should use a random one, and probe-request fingerprints can re-identify devices across address changes. They defeated randomisation for ~96% of Android phones and found a control-frame flaw affecting 100% of tested devices. The practical consequence is that probe sniffing is both biased (you see a non-random subset) and intrusive (you can re-identify that subset).

Cost. Low hardware cost: Musa and Eriksson used $70 access points. A modern monitor is €100–500. The compliance cost is high: legal review, DPIA, signage, retention, access control, and likely public consultation.

Privacy. MAC addresses are personal data (Autoriteit Persoonsgegevens, 2025). The processing is systematic monitoring of a publicly accessible area, which triggers a DPIA under Art. 35(3)(c) for large-scale deployments. The ePrivacy Directive's rules on location data and terminal equipment add another layer (ePrivacy Directive, 2002). The German supervisory authorities have repeatedly treated Wi-Fi tracking as personal-data processing (Unabhängiges Landeszentrum für Datenschutz Schleswig-Holstein, 2017), and the burden is on the controller to show necessity and proportionality. A city cannot rely on "we only count" if the raw data contain persistent identifiers and the purpose is commercial or operational analytics.

Deployment. The method is passive and cheap, which is why it persists. But it fails the two tests that matter for this paper: it does not measure people, and it cannot be made privacy-preserving by hashing alone. The only defensible use is a short, purpose-limited research study with ethics approval, a DPIA, immediate aggregation, and no cross-site linkage. That is not a footfall service.

Verdict. Not a solution. The paper should say so plainly and explain why the low hardware cost is a false economy.


2.6 Telco network data / mobile signalling

Principle. Mobile networks record when a device attaches, moves between cells, makes a call, or uses data. Call detail records (CDRs) and passive signalling data can be aggregated to estimate the number of subscribers in a cell or sector, and their flows between cells. The data are collected by the operator, not by the city.

Measured accuracy. Deville et al. (2014) used more than 1 billion call records from Portugal and France to produce national-scale population density estimates and compared them with census-based alternatives; they show that mobile-phone-based maps can reproduce population distributions at national scale, while stressing the need for calibration and anonymity. The spatial resolution is the limiting factor. The U.S. Bureau of Transportation Statistics puts cellular probe accuracy at 250–1,000 ft in urban areas, 1,500–3,000 ft in suburban areas and 1–3 miles in rural areas, and states that it is not precise enough for intersection-level or mode-separated analysis (BTS, 2025). Flowminder notes that urban subscribers are generally within about 500 m of a cell tower, with much lower density in rural areas (Flowminder, n.d.). A 2024 comparison of GPS, GSM and CDR for travel trajectories also finds higher measurement errors in GSM and CDR than in GPS, especially for departure times and travel distances (Zannat et al., 2024). That is fine for a district-level estimate and useless for a doorway count. Temporal resolution is also limited by when the network records an event; passive signalling is denser than CDR but still not a continuous people counter.

Cost. There is no sensor to buy. The cost is a data licence or a data-sharing agreement with one or more operators. These are bespoke: price depends on spatial and temporal resolution, history, exclusivity, and whether the operator does the aggregation. A planning range of €10k–100k+ per year for an aggregated city-level feed is realistic (est.), but it is an order of magnitude, not a quote. In the U.S., major carriers stopped sharing CDR data with analytics vendors in 2018 after privacy and public-relations concerns (BTS, 2025); the European market is different but still contract-driven.

Privacy. This is high-risk data. Even pseudonymised subscriber IDs can be re-identified through mobility patterns, home/work inference, and linkage with other data. The GDPR applies; the ePrivacy Directive applies to location data; a DPIA and a robust data-processing agreement are mandatory; aggregation to a minimum cell size and a minimum count threshold is necessary but not sufficient. The operator is usually the controller or joint controller for the raw data; the city should receive only aggregated, thresholded outputs and should not have access to subscriber-level records. The purpose must be specified and limited; "smart city analytics" is too broad.

Deployment. The city needs operator cooperation, a legal basis, a procurement route, and a data governance structure. Coverage and market share vary; prepaid users, children, tourists and multiple-SIM users distort the picture. The data are usually delivered with a lag (hours to days), which is fine for planning and useless for real-time staffing. The BTS guide also warns that the methodology for cellular probe data is dated in the U.S. because of the 2018 withdrawal; in Europe, ask the operator for a current validation.

Best use. City-wide context: district-level population, origin–destination flows, tourism and event impacts, and a prior for downscaling. It is not a street-level footfall sensor.

Verdict for Heilbronn. Useful as a city-wide covariate or prior, not as the measurement layer. The spatial resolution and governance burden rule it out for doorway or plaza counts.


2.7 GPS probe data / map providers

Principle. Apps and navigation services collect GPS traces from opted-in users. The provider map-matches the traces to a road or path network and sells aggregated counts, speeds or origin–destination matrices. Strava Metro is the best-known pedestrian/cyclist example; INRIX and TomTom are vehicle-oriented.

Measured accuracy. GPS itself is accurate to a few metres in open sky, but the sample is not the population. Venter et al. (2023) compared Strava counts with observed recreational use and found that Strava users are biased toward cyclists (+8%), males (+15.7%) and ages 35–54 (+20.4%). Walkers and hikers are underrepresented; pedestrian data are rare (Lee & Sener, 2020). Strava Metro's own case study reports model accuracy of ±32 to ±160 cyclists on 75% of streets — a wide interval that reflects the difficulty of scaling from a biased sample to total volumes (Strava Metro, n.d.). In indoor spaces, in urban canyons, and in dense historic centres, GPS drops out or degrades; the inner city of Heilbronn has exactly the kind of narrow streets and dense building frontage where this matters.

Cost. Strava Metro was priced at about $0.55 per app user per year in the NCDOT review (Ozan et al., 2021); a city with 130,000 residents and a 10% Strava user base would pay roughly $7,000 per year at that historical rate, but commercial pricing has changed and should be quoted. A commercial probe-data subscription is typically €10k–50k per year (est.). The hidden cost is validation: without independent counts, you cannot know the sampling fraction.

Privacy. GPS traces are among the most sensitive mobility data. Even aggregated and de-identified, they can reveal home, work, health-care visits and associations. The GDPR applies; consent and purpose limitation are central; commercial re-use and onward sharing need careful contracts. Strava's privacy settings and aggregation reduce but do not eliminate the risk.

Deployment. Coverage bias is the core problem: fit, male, middle-aged cyclists are overrepresented; children, older people, women, and people without smartphones are underrepresented. The data are good for route choice and recreational demand, weak for total footfall. Indoor and covered areas are invisible. Licensing and data-sharing agreements take time.

Best use. Cyclist route and origin–destination analysis, recreational demand, and long-term trend detection where the sample bias is stable and can be calibrated. Not inner-city pedestrian footfall.

Verdict for Heilbronn. Not suitable for the inner-city footfall question. It could be a supplementary dataset for cycling or tourism, clearly labelled as a biased sample.


2.8 Transit smart cards, ticket validators and turnstiles

Principle. Automated fare collection (AFC) records every validation: card or ticket ID, time, stop or station, and sometimes the vehicle. Turnstiles physically count entries. Automated passenger counters (APCs) on vehicles use infrared, camera or pressure sensors to count boardings and alightings. The data are administrative by-products, not a research sensor.

Measured accuracy. For the population that validates, the record is transaction-level: the validator either read the card or it did not. The problem is coverage. Smart card data capture only trips paid by smart card; cash passengers, non-validating passengers, and fare evaders are missing (Pelletier et al., 2011). Munizaga et al. (2020) estimate a 5% partial evasion rate for bus trip stages before a metro trip and a 37% total fare evasion rate for bus-only trips in their system. Caltrans (2025) reports that fare evasion in some California agencies rose from about 20% to nearly 50% after the pandemic, which directly biases ridership counts that rely on fare equipment. APCs can fill part of the gap, but they need their own validation against manual counts and against fare data.

Heilbronn's transit context is substantial: the Stadtwerke Heilbronn operates 14 city bus lines and the S4/S41/S42 Stadtbahn lines, with about 22.8 million passengers per year (5.5 million Stadtbahn, 17.3 million bus) (Stadtwerke Heilbronn, n.d.). That is a strong, high-frequency signal for commuting, shopping and event flows. It is not a pedestrian count for the Innenstadt, because transit passengers are a subset of pedestrians and the validators do not see people who walk or cycle.

Cost. If the AFC system already exists, the marginal cost of an aggregated data extract is low: a data-sharing agreement, an analyst, and a privacy review. New validators or turnstiles are capital projects with vehicle/station installation costs; a planning range of €5k–15k per vehicle or gate is plausible (est.), but it depends heavily on the system. APCs are cheaper than turnstiles but need calibration.

Privacy. Individual trip records are personal data. They are pseudonymous, not anonymous, and can be linked across days. The GDPR applies; the legal basis for the transit operator is usually the transport contract or a public task, but sharing with the city for analytics needs a separate basis and a DPIA. Aggregate to stop-level or line-level counts, apply a minimum count threshold, and do not publish trip-level data.

Deployment. The city needs the operator's cooperation and a clear data-sharing agreement. The data are usually available at stop/line level, with a delay; real-time feeds exist but are operational, not analytical. Fare policy and validation rules change the denominator. The transit network covers only part of the pedestrian system.

Best use. A high-quality covariate for transit demand, commuting patterns and event flows; a cross-check for city-wide models; not a general footfall sensor.

Verdict for Heilbronn. Use aggregated validator counts as a covariate in the fusion model. Do not use them as the denominator for inner-city footfall.


2.9 Manual counts and video annotation as ground truth

Principle. A trained observer stands at a cordon or screen line and counts people with a clicker or a tally sheet, usually by direction and sometimes by mode. Video annotation is the same task performed on recorded footage, which allows review and inter-rater checks.

Measured accuracy. Manual counting is accurate but not perfect. Diogenes et al. (2007) compared methods at intersections and found an average error of 11% for manual clickers, with 8% and 15% on two different dates. The error is systematic in one direction: the NCDOT review notes that manual entry is prone to undercounting and that error rates tend to be greater at the beginning of a session (Ozan et al., 2021). The TREC guide is blunt: a human can be expected to count accurately for about two hours; estimates of annual average daily pedestrian/bicycle traffic from just two hours of counts can have as much as 60% error on average; one week of counts substantially improves accuracy (Nordback, 2014). Video annotation removes weather and field fatigue but adds digitisation cost and its own inter-rater error; it is the standard way to audit automated counters.

Cost. In Germany, the 2026 minimum wage is €13.90 per hour (BMAS, 2026). A loaded observer rate — wage plus employer contributions, training, supervision, travel and equipment — is roughly €25–40 per hour (est.). A campaign at four sites, seven days, four two-hour shifts per day, two observers per shift is 448 observer-hours, or €11k–18k (est.). Video annotation is similar or slightly cheaper per hour but requires a secure viewing environment and blurring/retention controls. These are planning estimates; a procurement quote will differ.

Privacy. Count-only manual data are not personal data. Video annotation is different: the footage contains identifiable people, so the same GDPR/DPIA logic as camera counting applies. The mitigation is to record only what is needed, blur faces at ingest, restrict access, delete after annotation, and never publish footage.

Deployment. Manual counts need a sampling design, not just a person with a clicker. Choose cordons that capture the relevant flow, stratify by day type and time of day, rotate observers, take breaks every two hours, and measure inter-rater reliability on a subsample. Weather, holidays, events and school terms all affect the count; record them. For Heilbronn, a twice-yearly campaign at 3–4 sites, plus event-specific counts, is a realistic ground-truth program.

Best use. Calibration and validation of every automated sensor; estimation of detection probabilities; audit of vendor claims; occasional event evaluation. It is the only modality in this note that can claim to measure people directly.

Verdict for Heilbronn. Essential. A sensing stack without a manual/video calibration program is an unvalidated index. Budget for it from the start.


2.10 Sensor fusion / Bayesian hierarchical models

Principle. Each sensor measures a noisy, biased transformation of the same latent quantity: the number of people passing a site in a time interval. A fusion model treats the latent count as the parameter of interest and each sensor as an observation with its own detection probability, bias, and error structure. Hierarchical priors share information across sites and time, which is essential when calibration data are sparse.

Measured accuracy. There is no single accuracy number. The value of fusion is that it makes the uncertainty explicit and lets sparse ground truth calibrate many sensors. Parsons et al. (2022) show how a Bayesian hierarchical model can combine multiple estimates from different sources when none is complete; Zaouche and Bode (2023) formulate Bayesian spatio-temporal models to map urban pedestrian traffic from sparse count data. The transport literature has used hierarchical models for traffic counts and small-area estimation for decades; the same machinery applies here (Rao & Molina, 2015; Gelman et al., 2013). The practical test is not "does it beat the best sensor?" but "does it give a calibrated interval and a defensible estimate when the best sensor is missing?".

Cost. Software and engineering, not hardware: €20k–60k for model development, data pipeline and validation; €5k–15k per year for maintenance and re-calibration (est.). The biggest cost is the calibration program, which is covered under manual counts.

Privacy. The fused output is usually a count with an interval, which can be published as open data. The privacy risk comes from the inputs: if the model ingests camera frames, MAC addresses or individual traces, it inherits their risk. The design rule is to fuse aggregates and counts, not raw personal data.

Deployment. Fusion needs a data model, a calibration plan, and someone who can explain the output. It fails when the sensors are correlated in the same way (e.g., all undercount in rain) or when the calibration data are used both to fit and to evaluate. It also fails when the organisation wants a single number and ignores the interval. The model must be versioned, documented and re-estimated when sensor behaviour drifts.

Best use. The core of a multi-sensor city network: combine thermal, camera, radar, manual and transit data; downscale city-wide priors to street level; produce counts with credible intervals.

Verdict for Heilbronn. This is the "right way" layer. Buy fewer sensors and invest in the model and calibration.


3. Fusion and calibration

3.1 The latent-variable model

Let (N_t) be the true number of people crossing a site in interval (t) (e.g., 15 minutes). Sensor (s) produces (y_{st}). A defensible baseline is a negative-binomial observation model:

[ y_{st} \mid N_t, q_{st} ;\sim; \mathrm{NegBin}!\left(N_t q_{st},\ \phi_s\right) ]

where (q_{st} \in (0,1)) is the sensor's detection probability and (\phi_s) controls overdispersion. The latent count is modelled on the log scale with time, weather, holiday and event effects:

[ \log N_t = \alpha + f_{\text{time}}(t) + f_{\text{weather}}(t) + u_t, \qquad u_t \sim \mathrm{AR}(1) ]

The detection probability is modelled hierarchically, with partial pooling across sites:

[ \mathrm{logit}(q_{st}) = \beta_s + \gamma^\top x_t + \varepsilon_{st}, \qquad \beta_s \sim \mathcal{N}(\mu_\beta, \tau_\beta^2) ]

where (x_t) includes light, rain, temperature, crowding and sensor health. Manual counts (m_t) enter as a high-quality observation:

[ m_t \mid N_t ;\sim; \mathrm{NegBin}(N_t, \phi_m) ]

or as a lognormal with a known coefficient of variation. The posterior is proportional to the product of the sensor likelihoods, the manual likelihood, the latent-count prior and the detection-probability prior. Report the posterior median and an 80% credible interval for (N_t). The interval is the product; a point estimate without it is not a measurement.

This structure does three things that a simple average cannot. It separates the true count from sensor-specific detection, so a sensor that sees 40% of people is not treated as a sensor that sees 100%. It lets a well-calibrated site inform a poorly calibrated one through the hierarchical prior. And it propagates the uncertainty from the manual counts into the final estimate.

3.2 Calibration design

Calibration is the difference between a sensor network and a measurement system. A workable program for Heilbronn:

  1. Sites. Calibrate at 3–4 sites that represent the main geometries: a narrow doorway, a wide plaza, a transit stop, and a street segment.
  2. Ground truth. Two intensive one-week campaigns per year at 3–4 sites: four two-hour shifts per day (morning, midday, afternoon, evening), two observers per shift. That is 448 observer-hours per campaign at four sites. Between campaigns, run one-day checks at each site to track drift.
  3. Video audit. Record a subsample and annotate it independently. Use it to estimate inter-rater reliability and to check the field counts. Blur faces at ingest and delete the footage after annotation.
  4. Covariates. Record weather, light, events, holidays, and sensor health at each calibration interval.
  5. Model. Fit the hierarchical model with partial pooling. Estimate (q_{st}) by site, season and light condition. Check posterior predictive intervals against held-out counts.
  6. Drift. Track (q_{st}) over time with control charts. If a site's detection probability shifts by more than about 5 percentage points, inspect the sensor and re-calibrate.
  7. Validation. Hold out entire days, not random intervals, to avoid leakage. Report MAPE, bias, 80% interval coverage and rank correlation. A model that is accurate on average but has 50% interval coverage is not calibrated.

3.3 Identifiability, confounding and double counting

The model is only identifiable if at least one sensor is calibrated or if the priors are strong. If every sensor is an uncalibrated index, the data identify relative variation, not the level. This is the core problem with the existing PAX network: the device counts are informative about temporal patterns, but the mapping to people is unknown.

Three practical safeguards:

  • Negative controls. Use closed days, night hours and known-zero periods to estimate false positives. A sensor that counts people in an empty street is not calibrated.
  • Positive controls. Use events with independent counts (ticketed events, transit validations, manual cordons) to anchor the upper range.
  • Overlapping sensors. Where two sensors cover the same flow, compare their counts and estimate capture–recapture. Do not assume independence: sensors that share a failure mode (rain, darkness) will be correlated. Use a log-linear model with an interaction term, or treat the overlap as a diagnostic rather than a correction.

Double counting is a design problem, not a modelling problem. Define the cordon, the direction, and the time window before installation. A person who enters and leaves twice is two crossings; a person who stands in the zone is occupancy, not flow. Decide which quantity the decision needs: footfall (crossings), occupancy (people present), or unique visitors (deduplicated over a period). They require different sensors and different models.

3.4 Reporting and decision rules

The output should be three numbers, not one: a point estimate, an 80% interval, and a calibration date. For staffing, use the lower bound of the interval for the minimum staff calculation and the central estimate for the expected cost. If the interval width exceeds 50% of the median, label the series "trend only" and do not use it for staffing. If the holdout bias exceeds 10%, re-calibrate before publishing. These thresholds are design choices, not universal truths; state them in the paper and in the dashboard.

3.5 A minimal implementation plan

  1. Start with a calibration-factor model: one detection probability per sensor, estimated from manual counts, with bootstrap intervals. This is simple, transparent and enough for a pilot.
  2. Move to the hierarchical negative-binomial model when there are enough calibration days to estimate partial pooling.
  3. Fit in PyMC, Stan or a comparable probabilistic programming language; version the model and the data; publish the model card.
  4. Validate on held-out days and on a site that was not used for fitting.
  5. Publish the counts, intervals, calibration dates and code; do not publish raw video, MAC addresses or individual traces.

4. Recommendations for Heilbronn

4.1 Context

Heilbronn has about 131,986 residents and 100 km² of city area (Bitkom, 2026). The inner city is a compact retail and leisure cluster; the PulseCast data cover eight PAX sensors in a roughly 580 m × 280 m area, at 15-minute cadence, from June to November 2026. The values are device estimates, not people (PulseCast data notes, 2026). The city already has a digital strategy, a LoRaWAN initiative and an open-data programme (Stadt Heilbronn, n.d.-a, n.d.-b), which lowers the deployment cost for a low-power sensor network. The "Zukunft Innenstadt" programme already frames the inner city as a policy priority (Stadt Heilbronn, n.d.-c). The transit system carries about 22.8 million passengers per year (Stadtwerke Heilbronn, n.d.), a strong covariate for commuter and event flows.

The goal is not a city-wide heatmap. It is a defensible, privacy-preserving measurement layer for the inner city that can support staffing, event operations and urban planning, with uncertainty that is honest enough to make decisions on.

4.2 Design principles

  1. Measure people, not devices. Every sensor must have a documented relationship to a people count, calibrated locally. If it does not, label it an index.
  2. Use the least invasive sensor that meets the accuracy requirement. Thermal before camera; never face recognition; never MAC addresses.
  3. Process at the edge, store aggregates. No raw video, no raw point clouds, no MAC addresses. Counts and confidence only.
  4. Calibrate continuously. A regular manual/video campaign, at least twice a year, is part of the capital plan, not an afterthought.
  5. Publish uncertainty. Every count comes with an interval and a calibration date. The dashboard should not show a single number without context.
  6. Keep it open and vendor-neutral. Open data, open code, documented APIs, no lock-in to a single sensor vendor.
  7. Do the legal work first. DPIA, legal basis, signage, retention, processor agreements, and a public explanation before the first sensor is installed.

4.3 Reference stack

Tier Sites Modality Output Expected accuracy Role
1. Permanent core 8–12 narrow gateways: store entrances, passages at Stadtgalerie, Marahaus, Götzenturm, Soleo Low-resolution thermal array (8×8 or 32×24), on-device counting, LoRaWAN/Wi-Fi backhaul 15-min in/out counts, occupancy 90–95% per passage under good conditions; lower in heat, groups or wide openings Default measurement layer; privacy-first; low cost
2. Reference / wide areas 2–4 wide plazas or complex crossings: Hagenbuchersee, Neckarufer, event spaces Edge AI camera, on-device inference, no face recognition, frames discarded Counts, direction, pedestrian/cyclist classification 10–20% MAPE expected outdoors; validate locally Accuracy benchmark; wide-area coverage; validation
3. Gap-fill 1–2 weather-exposed or wide crossings, transit stops 60 GHz mmWave (and optionally one LiDAR reference) Presence, flow, count with wide interval 69–96% presence/count depending on scene; ±25–50% for small crowds Cover where thermal/camera cannot; not for exact staffing
4. Ground truth 3–4 representative sites, twice yearly Manual counts + video annotation Cordon counts by direction and time ~11% manual error; improve with training and video audit Calibration, validation, audit
5. Context City-wide Transit validators, weather, events, holidays; optionally telco/GPS aggregates Covariates and priors Transit exact for validators; telco 250–1,000 ft urban; GPS biased Explain demand; downscale city-wide trends

Site logic. Use thermal where the passage is narrow and the flow is single-file or two-file. Use cameras where the area is wide, the count is high-value, and the privacy assessment supports it. Use radar where the width or weather rules out thermal and the budget rules out camera. Do not put a thermal array on an open plaza and expect 90% accuracy; the geometry is wrong.

A note on the existing PAX sensors. Keep two or three in place for continuity during the transition. Run them side-by-side with the new sensors for 12 weeks, estimate a calibration ratio, and publish it. Then either phase them out or keep them as a temporal covariate, clearly labelled as a device index. Do not present the historical series as people; the penetration rate is unknown and has probably changed over five months.

4.4 Pilot protocol

A 12-week pilot at six sites is enough to estimate accuracy and operational cost without committing to a city-wide rollout.

Sites. Two narrow doorways (thermal), two wide plazas (camera), one transit stop (radar), one control site with manual counts only. Use the existing PAX locations where possible to allow side-by-side comparison.

Ground truth. Two intensive one-week campaigns (weeks 1 and 12), four two-hour shifts per day, two observers per shift, rotating across 3–4 sites; 448 observer-hours per campaign at four sites. Video a subsample for audit. Record weather, events, holidays and sensor health.

Metrics. MAPE and bias against manual counts; 80% interval coverage; correlation with the PAX index; availability (percentage of expected 15-minute intervals with data); cost per site-year; privacy incidents; installation time.

Go/no-go thresholds. A sensor is fit for staffing if MAPE ≤ 20%, absolute bias ≤ 10%, 80% interval coverage ≥ 75%, availability ≥ 95%, and there are no unresolved privacy issues. If it misses, use it for trends only. These thresholds are deliberately modest; a 20% error is still useful for a weekly shift plan, but it is not a headcount.

Analysis. Fit the hierarchical model on the first eight weeks and validate on the last four. Publish the pre-registered metrics and the model card. Do not tune the model on the validation weeks.

  • Legal basis. For a city, Art. 6(1)(e) GDPR (public task) is the likely basis for footfall measurement; document the task and the necessity. For a private operator, Art. 6(1)(f) may apply, but a public-space deployment needs a careful balancing test.
  • DPIA. Required for systematic monitoring of a publicly accessible area on a large scale (Art. 35(3)(c) GDPR). Do it before procurement, not after. Include a necessity and proportionality assessment, a data-flow diagram, a retention schedule, and a risk register.
  • BDSG §4. Video surveillance of publicly accessible spaces must be necessary for the task, the purpose must be specified, and data must be deleted when no longer needed (BDSG, 2017). Apply the same logic to any modality that can identify people.
  • AI Act. No face recognition, no biometric categorisation, no emotion recognition. People counting without identification is outside the Art. 5(1)(h) prohibition, but adding identification would change the legal analysis (AI Act, 2024).
  • Transparency. Signage at each site, a public web page explaining what is measured, why, how long it is kept, and how to complain. For thermal, say explicitly that no images are formed.
  • Retention. Zero raw frames; counts retained indefinitely as open data; calibration video deleted after annotation; access logs retained for audit.
  • Procurement. Require on-device processing, no hidden cloud upload, open APIs, data ownership by the city, security testing, and an exit/portability clause.
  • Data protection by design. Edge inference, encryption, signed firmware, least privilege, no default passwords, regular updates, incident response.
  • Publication. Open 15-minute counts with intervals and calibration dates; a data dictionary; a DPIA summary; the model code. Do not publish raw video, MAC addresses or individual traces.

4.6 Budget (planning estimate)

Item Year-1 range (€) Basis
8–12 thermal nodes 2,000–5,000 €150–400 per node; sensor ~$50
2–4 edge cameras 1,500–8,000 €500–2,000 per node; commercial systems higher
1–2 mmWave nodes 600–1,400 €300–700 per node
Installation, mounting, power, backhaul 8,000–20,000 Site-dependent; assumes use of existing street furniture and LoRaWAN/Wi-Fi
Backend, dashboard, integration 20,000–40,000 Data pipeline, API, open-data portal, security
Calibration (manual + video) 22,000–36,000 Two campaigns per year, ~448 observer-hours each; ~€25–40 per observer-hour
DPIA, legal, procurement 8,000–15,000 External counsel and DPO time
Project management, training, comms 8,000–15,000 Part-time coordination, training and public communication
Contingency (15%) 10,500–21,000 Hardware delays, site changes
Total, year 1 ~81,000–161,000 All figures are estimates, not quotes
Recurring, year 2+ 30,000–50,000 Calibration, maintenance, cloud, replacements

The budget is deliberately hardware-light and calibration-heavy. That is the right ratio. A network of cheap uncalibrated sensors is not cheaper than a smaller network with a real ground-truth program; it is just less honest.

4.7 What not to do

  • Do not use BLE or Wi-Fi MAC sniffing for public footfall. The data are biased, the penetration is unknown, and the privacy cost is high. The paper's critique is correct.
  • Do not buy a "95% accuracy" claim without a local validation. The same sensor can be 95% in a controlled doorway and 70% on a windy plaza. Ask for the evaluation protocol, the ground truth and the error distribution.
  • Do not use a single sensor for staffing. Every modality has a failure mode; fuse at least two, and calibrate against manual counts.
  • Do not store raw video or point clouds. The marginal analytical value is small; the legal and reputational risk is large.
  • Do not use face recognition, gait recognition or re-identification. It is unnecessary for counting and changes the legal regime.
  • Do not compare counts across sensors without calibration. A thermal count, a camera count and a radar count are different measurements until they are mapped to the same latent quantity.
  • Do not present device counts as people. Label the historical PAX series as an index and calibrate it or retire it.

5. Citations

Ahamed, I., Ranathunga, C. D., Udayantha, D. S., Ng, B. K. K., & Yuen, C. (2024). Real-Time AI-Driven People Tracking and Counting Using Overhead Cameras. arXiv:2411.10072. https://arxiv.org/abs/2411.10072

Autoriteit Persoonsgegevens (Dutch Data Protection Authority). (2025). Wifi and bluetooth. https://www.autoriteitpersoonsgegevens.nl/en/themes/internet-and-smart-devices/wifi-and-bluetooth

Bitkom. (2026). Smart City Index 2026: Heilbronn. https://www.bitkom.org/Smart-City-2026/Heilbronn

Brena, R. F., Escudero, E., Vargas-Rosales, C., Galvan-Tejada, C. E., & Munoz, D. (2021). Device-free crowd counting using multi-link Wi-Fi CSI descriptors in Doppler spectrum. Electronics, 10(3), 315. https://doi.org/10.3390/electronics10030315

Bundesdatenschutzgesetz (BDSG). (2017). § 4 Videoüberwachung öffentlich zugänglicher Räume. https://www.gesetze-im-internet.de/bdsg_2018/__4.html

Bundesministerium für Arbeit und Soziales (BMAS). (2026). Allgemeiner gesetzlicher Mindestlohn. https://www.bmas.de/DE/Arbeit/Arbeitsrecht/Mindestlohn/mindestlohn.html

California Department of Transportation (Caltrans). (2025). PI-0393 Fare Evasion and its Impact on Measured Ridership (T. Monson, author). https://dot.ca.gov/-/media/dot-media/programs/research-innovation-system-information/documents/preliminary-investigations/pi-0393-fare-evasion-and-effects-on-ridership-counts-final-v2-a11y.pdf

Cyganek, B., Janyst, P., Knapik, M., Przebinda, Ł., Ptaszkiewicz, M., & Tajmajer, T. (2022). Evaluation of an Embedded Computer Vision System for Outdoor Media Audience Analysis. PP-RAI 2022. https://wydawnictwo.umg.edu.pl/pp-rai2022/pdfs/07_pp-rai-2022-101.pdf

Deville, P., Linard, C., Martin, S., Gilbert, M., Stevens, F. R., Gaughan, A. E., Blondel, V. D., & Tatem, A. J. (2014). Dynamic population mapping using mobile phone data. Proceedings of the National Academy of Sciences, 111(45), 15888–15893. https://doi.org/10.1073/pnas.1408439111

Diogenes, M. C., Greene-Roesel, R., Arnold, L. S., & Ragland, D. R. (2007). Pedestrian counting methods at intersections: A comparative study. Transportation Research Record, 2002(1), 26–30. https://doi.org/10.3141/2002-04

European Data Protection Board (EDPB). (2020). Guidelines 3/2019 on processing of personal data through video devices (Version 2.0). https://www.edpb.europa.eu/sites/default/files/files/file1/edpb_guidelines_201903_video_devices_en_0.pdf

European Union. (2002). Directive 2002/58/EC of the European Parliament and of the Council of 12 July 2002 concerning the processing of personal data and the protection of privacy in the electronic communications sector (ePrivacy Directive). https://eur-lex.europa.eu/eli/dir/2002/58/oj

European Union. (2016). Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 (General Data Protection Regulation). https://eur-lex.europa.eu/eli/reg/2016/679/oj

European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act). https://eur-lex.europa.eu/eli/reg/2024/1689/oj

Flowminder. (n.d.). Strengths and limitations of Call Detail Records (CDR data). https://flowgeek.org/what-are-cdrs/what-are-the-strengths-and-limitations-of-cdr-data

Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., & Rubin, D. B. (2013). Bayesian Data Analysis (3rd ed.). CRC Press.

Kurkcu, A., & Ozbay, K. (2017). Estimating pedestrian densities, wait times, and flows with Wi-Fi and Bluetooth sensors. Transportation Research Record, 2644(1), 72–82. https://doi.org/10.3141/2644-09

Lee, K., & Sener, I. N. (2020). Emerging data for pedestrian and bicycle monitoring: Sources and applications. Transportation Research Interdisciplinary Perspectives, 4, 100095. https://doi.org/10.1016/j.trip.2020.100095

Lesani, A., Nateghinia, E., & Miranda-Moreno, L. F. (2020). Development and evaluation of a real-time pedestrian counting system for high-volume conditions based on 2D LiDAR. Transportation Research Part C: Emerging Technologies, 114, 20–35. https://doi.org/10.1016/j.trc.2020.01.018

Maaspuro, M. (2020). A low-resolution IR-array as a doorway occupancy counter in a smart building. International Journal of Online and Biomedical Engineering, 16(6), 4–18. https://doi.org/10.3991/ijoe.v16i06.13915

Martin, J., Mayberry, T., Donahue, C., Foppe, L., Brown, L., Riggins, C., & Rye, E. C. (2017). A study of MAC address randomization in mobile devices and when it fails. Proceedings on Privacy Enhancing Technologies, 2017(4), 365–383. https://doi.org/10.1515/popets-2017-0054

Mohammadmoradi, H., Munir, S., Gnawali, O., & Shelton, C. (2017). Measuring people-flow through doorways using easy-to-install IR array sensors. Proceedings of the 13th International Conference on Distributed Computing in Sensor Systems (DCOSS), 35–43. https://doi.org/10.1109/DCOSS.2017.26

Munizaga, M. A., Gschwender, A., & Gallegos, N. (2020). Fare evasion correction for smartcard-based origin-destination matrices. Transportation Research Part A: Policy and Practice, 141, 307–322. https://doi.org/10.1016/j.tra.2020.09.008

Musa, A. B. M., & Eriksson, J. (2012). Tracking unmodified smartphones using Wi-Fi monitors. Proceedings of the 10th ACM Conference on Embedded Network Sensor Systems (SenSys), 281–294. https://doi.org/10.1145/2426656.2426685

Nordback, K. (2014). Guide to Bicycle & Pedestrian Count Programs. Transportation Research and Education Center, Portland State University. https://trec.pdx.edu/sites/default/files/smol-IBPI%20Guide%20to%20Bicycle%20%26%20Pedestrian%20Count%20Programs_final.pdf

Ozan, E., Searcy, S., Geiger, B. C., Vaughan, C., Carnes, C., Baird, C., & Hipp, A. (2021). State-of-the-Art Approaches to Bicycle and Pedestrian Counters (NCDOT Report No. FHWA/NC/2020-39). https://connect.ncdot.gov/projects/research/RNAProjDocs/RP2020-39%20Final%20Report.pdf

Pallaprolu, A., Peng, P., Sandhu, S., Hurst, W., & Mostofi, Y. (2024). Crowd analytics with a single mmWave radar. Proceedings of the 30th Annual International Conference on Mobile Computing and Networking (MobiCom), 830–844. https://doi.org/10.1145/3636534.3690664

Parsons, J., Niu, X. M., & Bao, L. (2022). A Bayesian hierarchical model for combining multiple data sources in population size estimation. Annals of Applied Statistics, 16(3), 1550–1562. https://doi.org/10.1214/21-AOAS1556

Pelletier, M.-P., Trépanier, M., & Morency, C. (2011). Smart card data use in public transit: A literature review. Transportation Research Part C: Emerging Technologies, 19(4), 557–568. https://doi.org/10.1016/j.trc.2010.12.003

Perra, C., Kumar, A., Losito, M., Pirino, P., Moradpour, M., & Gatto, G. (2021). Monitoring indoor people presence in buildings using low-cost infrared sensor array in doorways. Sensors, 21(12), 4062. https://doi.org/10.3390/s21124062

PulseCast data notes. (2026). Data findings — PAX counters, Heilbronn. Internal project document: OKF_Vault/mission/data.md.

Rao, J. N. K., & Molina, I. (2015). Small Area Estimation (2nd ed.). Wiley.

Ren, L., Yarovoy, A. G., & Fioranelli, F. (2023). Grouped people counting using mm-wave FMCW MIMO radar. IEEE Internet of Things Journal, 10(22), 20107–20119. https://doi.org/10.1109/JIOT.2023.3282797

Sakhnini, A., Rykunov, M., Bourdoux, A., Pollin, S., & Berkvens, R. (2024). Crowd counting measurements in a festival area using a mmWave FMCW radar. Proceedings of the IEEE Radar Conference (RadarConf24), 1–5. https://doi.org/10.1109/RadarConf2458775.2024.10549396

Santos, J. D. D., Alarcon, R. J. F., Asuncion, K. F., Galang, C. J. S., Oliveros, H. G. V., & Reyes, J. M. R. (2024). Investigation of environment dependence in Wi-Fi CSI-based crowd counting systems. Proceedings of the IEEE Region 10 Conference (TENCON), 362–365. https://doi.org/10.1109/TENCON61640.2024.10903043

Stadt Heilbronn. (n.d.-a). Digitalisierung | KI. https://www.heilbronn.de/wirtschaft/digitalisierung.html

Stadt Heilbronn. (n.d.-b). LoRaWAN. https://www.heilbronn.de/wirtschaft/digitalisierung/digitale-infrastruktur/lorawan.html

Stadt Heilbronn. (n.d.-c). Zukunft Innenstadt Heilbronn. https://www.heilbronn.de/wirtschaft/zukunft-innenstadt-heilbronn.html

Stadtwerke Heilbronn. (n.d.). Bus & Bahn. https://www.stadtwerke-heilbronn.de/swh/bus-und-bahn/

Strava Metro. (n.d.). Using Strava data for active transportation planning. https://metro.strava.com/case-studies/using-strava-data-for-active-transportation-planning

Umezu, N., Koizumi, S., Nakagawa, K., & Nishida, S. (2023). Potential of low-cost light detection and ranging (LiDAR) sensors: Case studies for enhancing visitor experience at a science museum. Electronics, 12(15), 3351. https://doi.org/10.3390/electronics12153351

U.S. Department of Transportation, Bureau of Transportation Statistics. (2025). BTS Transportation Probe Data Guide: Non-LBS Cellular Probe Data. https://rosap.ntl.bts.gov/view/dot/90548

Unabhängiges Landeszentrum für Datenschutz Schleswig-Holstein (ULD). (2017). Tätigkeitsbericht 39, Kapitel 7: Neue Medien. https://www.datenschutzzentrum.de/tb/tb39/kap07.html

Vaidya, G., & Zúñiga, M. A. (2024). Exploiting mmWave and deep-learning models to estimate people count in urban scenarios. Proceedings of the 9th IEEE/ACM International Conference on Internet-of-Things Design and Implementation (IoTDI), 73–84. https://doi.org/10.1109/IoTDI61053.2024.00011

Venter, Z. S., Gundersen, V., Scott, S., & Barton, D. N. (2023). Bias and precision of crowdsourced recreational activity data from Strava. Landscape and Urban Planning, 232, 104686. https://doi.org/10.1016/j.landurbplan.2023.104686

Wang, W., Vaidya, G., Bhattacharjee, A. K., Fioranelli, F., & Zúñiga, M. A. (2023). A long-term study of mmWave sensing in an outdoor urban scenario. Proceedings of the International Conference on Distributed Computing in Smart Systems and the Internet of Things (DCOSS-IoT), 240–247. https://doi.org/10.1109/DCOSS-IoT58021.2023.00049

Yonekura, H., Rizk, H., & Yamaguchi, H. (2025). Round trip time meets transformers: High-fidelity human counting in cluttered environments. Neural Computing and Applications, 37(28), 23591–23617. https://doi.org/10.1007/s00521-025-11540-8

Zannat, K. E., Choudhury, C. F., Hess, S., & Watling, D. P. (2024). Investigating the relative accuracy of GPS, GSM and CDR data for inferring spatiotemporal travel trajectories. IET Intelligent Transport Systems, 18(S1), 3013–3033. https://doi.org/10.1049/itr2.12563

Zaouche, M., & Bode, N. W. F. (2023). Bayesian spatio-temporal models for mapping urban pedestrian traffic. Journal of Transport Geography, 111, 103647. https://doi.org/10.1016/j.jtrangeo.2023.103647

Zhou, R., Lu, X. Y., Fu, Y. L., & Tang, M. (2020). Device-free crowd counting with WiFi channel state information and deep neural networks. Wireless Networks, 26(5), 3495–3506. https://doi.org/10.1007/s11276-020-02274-7


Appendix A. Evidence gaps to close locally

The literature gives ranges, not a Heilbronn answer. The pilot should answer these questions:

  1. What is the actual detection probability of a thermal array at each Heilbronn site, by season, light and crowd density?
  2. How does the existing PAX index correlate with manual counts, and is the relationship stable over the five-month window?
  3. What is the false-positive rate of mmWave at Hagenbuchersee and Neckarufer in rain and wind?
  4. How much does the camera's error rise at night and in the Christmas-market crowd?
  5. What is the cost per site-year, including calibration, maintenance and replacement?
  6. Does the fused estimate achieve 80% interval coverage on held-out days, or is it overconfident?
  7. Can the city publish the counts as open data without creating a re-identification risk, and what aggregation threshold is needed?

Until those questions are answered, the correct claim is "estimated people, with a calibrated interval," not "exact headcount."