Files

61 KiB
Raw Permalink Blame History

BLE PAX counters: how they count, how they fail, and what can be validated

Research note 01 — PulseCast / Future City Hackathon 2026 Prepared: 14 November 2026 Scope: passive BLE scanning for pedestrian counting; the 8-sensor Heilbronn network; implications for a privacy-first alternative.

Confidence labels used throughout:

Label Meaning
[Established] Replicated peer-reviewed finding or protocol fact.
[Primary-code] Verified directly in the public Code for Heilbronn firmware / commit.
[Forensic] Derived from the delivered CSV; reproducible with the code in Appendix A.
[Peer-reviewed] Published in a peer-reviewed venue; single study unless stated.
[Gray] Practitioner, government, vendor or thesis source; useful but not peer-reviewed.
[Inference] Reasoned from evidence; not directly measured.
[Uncertain] Conflicting or missing evidence.

The single most important correction. The PulseCast series is not a 15-minute count of people, and not a 15-minute aggregate of device counts. In the public Code for Heilbronn firmware, each sensor wakes every 900 s, scans BLE for about 10 s, counts distinct locally-administered advertising addresses in that window, multiplies the integer by 0.7, and transmits one value. The 15-minute cadence is the sampling interval, not the integration window. Every downstream claim about "people per 15 minutes" inherits this.

Caveat on firmware versions. The 2026 production window is exactly consistent with the public 2026 firmware: 100% of its 83,156 published values are exact multiples of 0.7. The older density-01 hardware produced integer values from October 2025 to March 2026, then switched to multiples of 0.7 in April 2026 — a firmware/calibration break. The exact scan duration of the 2025 build is not public. The direction of every argument below is unchanged whether the scan was 10 s or the configured 30 s.


Executive summary

  1. The unit is a short snapshot, not a 15-minute count. The public firmware scans for ≈10 s every 900 s — a temporal duty cycle of ≈1.1% (at most 3.3% if the configured 30 s were honoured). A passer-by is detected only if their dwell overlaps the scan window; a person who lingers is counted again in every 15-minute bin. The metric therefore conflates flow (people passing) with dwell (people staying), and is not a screenline crossing count. [Primary-code]

  2. The published value is 0.7 × U, where U is an integer count of distinct locally-administered BLE addresses in one scan. In the 8-sensor production window (2026-06-12 → 2026-11-07), 100% of 83,156 values are exact multiples of 0.7; the reconstructed U ranges 0–418, median 12, p95 71, p99 114, with 2.7% zeros. The older density-01 series is integer-valued until 2026-03, then switches to 0.7 — evidence of a firmware/calibration change, not a change in footfall. [Forensic]

  3. A single scalar cannot correct three different biases. The factor 0.7 would have to absorb (i) the share of people carrying any detectable advertising device (d < 1), (ii) the mean number of distinct advertising devices per carrier (m > 1), and (iii) the probability that a device advertises within the short scan (q < 1). The true relationship is y/P = 0.7 · d · m · q. No calibration data are published, and the default is a provisioning fallback, not a validated penetration correction. [Primary-code] [Inference]

  4. MAC randomization does not inflate the count within a 10-s scan, but it destroys everything across scans. RPA rotation is 900 s by default (range 1–3600 s), and measured address lifetimes are ~15–45 min. Within a 10-s window, rotation is rare; across 15-minute bins, the sensor's dedup bitmap is reset at every wake, so the same person is counted again. Randomization mainly prevents cross-sensor deduplication, dwell estimation and re-identification, not within-scan counting. [Established] [Primary-code]

  5. The counter silently excludes public-address devices and silently includes non-phone gadgets. The firmware counts only addresses with the locally-administered bit set. A phone or wearable using a public/static address is invisible; a smart TV, beacon, earbud case or car head-unit with a random address is counted. Vendor OUI cannot be used to separate phones from infrastructure because the counted addresses are random. [Primary-code] [Inference]

  6. The literature does not support absolute counts from BLE alone. Classic Bluetooth discoverability was 2–7% of pedestrians; modern BLE representativity in transit settings was 14–49%, with strong age bias. Absolute indoor count errors range from MAE 1.4–5.5 people (MAPE 24–91%, worst in low occupancy), and a BLE-only ridership model can have negative R² until fused with other data. The consistent finding is: relative trends yes, absolute levels no — unless calibrated against ground truth. [Peer-reviewed] [Gray]

  7. Environmental failure modes are large relative to the signal. Outdoor BLE range was 20–25 m for two phones in a vineyard test near Heilbronn; a human body adds 10–20 dB path loss (10 dB back-shadowing, 3 dB side at 3 m); Wi-Fi coexistence cut BLE throughput by ~41% in a shielded-chamber experiment; and LoRaWAN adds 1% duty-cycle constraints and real packet loss. Crowding adds collisions and queue drops. [Peer-reviewed] [Gray]

  8. The 15-minute series is heterogeneous and incomplete. Seven sensors have a median gap of 15.3–15.4 min; density-01 runs at 16.1 min. For the seven named sensors, 2.6–16.2% of intervals exceed 16 min; density-01 is at a 16.1-min median, so 99.8% of its intervals exceed 16 min. Outages reach ~2.5 h. Missingness is caused by LoRaWAN joins, transmission failures and power, and is not a random sample of footfall. [Forensic]

  9. The privacy claim is a design claim, not a property of the data. The current firmware reduces addresses to a 16-bit ID (the last two bytes; the blog calls this "hashing"), resets every wake, and transmits only counts — genuine data minimisation. But the sensing is still involuntary, the sensor is capable of raw capture, and the legal status of MAC-based counting in Germany is contested: one state DPA concluded that WLAN tracking for visitor counting was unlawful, and the DSK treats reading MAC addresses as touching terminal-equipment privacy. "Privacy-preserving" must be argued from architecture and governance, not asserted. [Gray] [Uncertain]

  10. Implication for PulseCast. Keep the forecaster, change the claim. Treat the series as a relative activity index in device-address units, validate it against manual counts, publish the unit and the factor, propagate calibration uncertainty into staffing decisions, and never present device counts as people. The privacy-first alternative must beat this sampling design — not a strawman — on cost, coverage, accuracy and privacy.


1. How BLE PAX counting works

1.1 Radio basics

BLE operates in the 2.4 GHz ISM band, divided into 40 channels of 2 MHz. Three channels are reserved for advertising: 37 (2402 MHz), 38 (2426 MHz), 39 (2480 MHz). Legacy advertising transmits the same payload on up to three of these channels, one at a time, with a random delay of 0–10 ms per advertising event; the advertising interval is configurable from 20 ms to 10.24 s. A passive scanner tunes to the advertising channels, reports the advertiser's 48-bit address and an RSSI value, and never needs to connect. [1][3]

Two consequences matter for counting:

  • The address is self-asserted. Before a connection is authenticated, the advertising address is whatever the device chooses. It can be public, random static, resolvable private (RPA) or non-resolvable private (NRPA). A listener cannot tell whether two addresses belong to the same physical device, or whether one device is impersonating another. [1][4]
  • RSSI is a noisy proxy for proximity, not distance. It has 1 dB resolution and is affected by orientation, body shadowing, multipath and channel. Using it as a hard range gate creates a detection zone that is neither circular nor stable. [20][21]

1.2 The PulseCast / Code for Heilbronn implementation

The Heilbronn network is maintained by Code for Heilbronn (CFHN). Their public blog and repository identify the hardware and firmware; the Fraunhofer IAO/KODIS project page documents the wider Heilbronn and Schwäbisch Hall deployment. [8]

  • Hardware: Heltec Wireless Paper / Heltec WiFi LoRa 32 V3 — ESP32-S3, SX1262 LoRa transceiver, BLE 4.2. [6][7]
  • Firmware: PlatformIO project sensor-pax, built on a vendored paxlib derived from the Deutsche Bahn libpax / ESP32-Paxcounter lineage. BLE only; Wi-Fi counting is compiled out. [7]
  • Counting rule: mac_add() accepts an advertising report only if the locally administered bit (paddr[0] & 0b10) is set; it then uses the last two bytes of the address as a 16-bit ID in a bitmap. New bit → increment; existing bit → ignore. The bitmap is cleared on every boot. [7] The blog calls this "hashing"; the code does not hash — it truncates the address to its low 16 bits. That is a dedup key, not a cryptographic hash, and it leaks the low bits of the address.
  • Scan parameters: passive scan; BLESCANWINDOW = 80 ms, BLESCANINTERVAL = 80 ms → 100% duty cycle during the scan; RSSI threshold -80 dBm; report interval 10 s; counter mode 1 (cumulative). [7]
  • Wake cycle: sleep_sec default 900 s; the node wakes, scans, waits for the first 10-s report callback, stops the scan, sends one LoRaWAN uplink, and deep-sleeps. [7]
  • Payload: paxCount * factor, with factor default 0.7 in provision.h; the value is formatted with four decimals. [7]

A code-level trace of one cycle:

wake (deep-sleep reset)
  └─ paxSetup(): blescantime=30, window=80ms, interval=80ms, RSSI≥−80 dBm
       libpax_counter_init(callback, count, 10 s, mode=1)
  └─ start_BLE_scan(): passive; duplicate filter disabled
  └─ first 10-s timer callback → new_data_available = true
  └─ paxStop(); paxCount = count_from_libpax.pax
  └─ payload = sprintf("%s,0,%.4f,1,%.4f", id, paxCount * cfg.factor, battery)
  └─ LoRaWAN TX; deep sleep 900 s

Why ≈10 s, not 30 s. paxSetup() sets configuration.blescantime = 30, but the vendored start_BLE_scan() accepts blescantime and never uses it: it only programs scan window and interval, then starts scanning. The only stop condition is main.cpp's wait for new_data_available, which is set by the 10-s report callback. The effective scan is therefore ≈10 s (plus up to 100 ms of loop delay). If a deployed build implemented blescantime as a hard stop, the scan would be 30 s and the duty cycle 3.3%; the argument below is unchanged in direction. [Primary-code, high confidence for the public commit; moderate confidence for the deployed 2025 binary]

Why the bitmap reset matters. With counter_mode = 1, report() does not reset the bitmap. But each wake is a cold boot from deep sleep, so the bitmap starts empty anyway. The dedup window is therefore one scan, not one hour and not one day. The 2025 blog says addresses are "hashed and stored for at most one hour" — that may describe an earlier firmware or a backend design; the public 2026 firmware does not implement a one-hour store. This is an open question (Section 7). [6][7]

1.3 What one published value means

Let:

  • U = number of distinct 16-bit IDs derived from locally-administered advertising addresses observed during one ≈10 s scan;
  • f = configured multiplier (default 0.7);
  • y = published "density" value.

Then:

y = f · U

U is not the number of people in the area, the number of devices in the area, the number of devices passing in 15 minutes, or the number of 15-minute unique visitors. It is a short-window count of distinct randomized advertising addresses, scaled by a constant.

If a person carries m advertising devices, is detectable with probability d (device ownership × Bluetooth state × advertising state), and their devices advertise within the scan with probability q, then the expected relationship to the true number of people present P is:

E[y] = f · d · m · q · P

The factor f = 0.7 can only be a correct people-per-device conversion if d · m · q = 1/0.7 ≈ 1.43. That is possible if d is high and m ≈ 1.4, but it is not measured. Assuming q ≈ 1 (devices in range for the full scan), if d = 0.5 and m = 1.4, y/P ≈ 0.49 — a 2× undercount. If d = 0.9 and m = 1.2, y/P ≈ 0.76 — a 24% undercount. The scale is therefore arbitrary without ground truth, and a single global factor cannot track seasonal, demographic or device-mix changes. [Inference]

1.4 Forensic evidence from the delivered CSV

The delivered file data/raw/schwarz-pax.csv contains 192,464 rows; 96,232 are densityData. The 8-sensor production window (2026-06-12 → 2026-11-07) contains 83,156 density rows across eight sensor names. Every one of those values is an exact multiple of 0.7:

Sensor (production window) Rows Values exact multiples of 0.7 Reconstructed U: median p95 max U = 0
hn-stadtgalerie-01 8,982 100% 12 96 196 0.0%
hn-neckarufer-01 9,527 100% 17 88 365 0.1%
hn-marahaus-01 9,254 100% 13 55 278 0.0%
hn-hagenbuchersee-02 10,099 100% 16 92 287 0.7%
hn-hagenbuchersee-01 12,345 100% 9 71 418 5.6%
hn-soleo-01 10,327 100% 12 40 159 4.8%
hn-goetzentuem-01 9,952 100% 12 42 301 3.0%
density-01 (f8c19730) 12,670 100% 11 40 90 5.5%
All eight 83,156 100% 12 71 418 2.7%

Interpretation. A continuous-valued physical series would not land on exact multiples of 0.7 in 83,156 consecutive observations. The only plausible explanation is that the published value is round(0.7 × integer, 4). The reconstructed U is an integer count of distinct randomized addresses in a scan. This is a direct, data-internal confirmation of the firmware reading. [Forensic]

The calibration break. The older density-01 hardware (863f75b0, 2025-10-16 → 2026-03-12) is 100% integer-valued and has a mean of 1.7–5.5 depending on month. The replacement hardware (f8c19730, 2026-04-16 → 2026-11-07) is 100% multiples of 0.7 and has a mean of 8.3–11.6 across full months. The series therefore splices two instrument regimes: an integer-count firmware and a 0.7-scaled firmware. Any model that uses the 13-month density-01 history across the break is fitting a calibration change, not footfall. [Forensic]

What this does not prove. The data cannot tell us the deployed scan duration, the exact factor value per sensor (only that 0.7 appears throughout the delivered 2026 production data), whether the 2025 firmware used a one-hour dedup window, or whether the sensors were re-provisioned. Those are open questions (Section 7).

1.5 Dedup, dwell time and the 15-minute cadence

A PAX counter can estimate dwell by tracking a persistent ID across consecutive scans and recording first/last seen. This one does not:

  • Within a scan, the bitmap deduplicates by a 16-bit ID taken from the last two address bytes. Two devices sharing those bytes collide (expected collisions ≈ U(U−1)/(2·65536); negligible at U = 12, ~1.3 at U = 418). A device that changes only its low bytes is counted as new.
  • Across scans, the bitmap is gone. There is no first/last-seen timestamp in the payload. The only data that leave the sensor are U, U × 0.7 and battery voltage.
  • Dwell resolution would be at best 15 min even if IDs persisted, and RPA rotation (~15 min default) caps observable track length. Any dwell estimate beyond ~15 min is structurally impossible with this design.

What repeated counts can proxy is aggregate persistence: a person who stays for an hour is counted in up to four bins, while a passer-by is counted with probability ≈ (scan + dwell)/900. But this is confounded with arrival rate and with the number of devices per person, and cannot be inverted without strong assumptions. Do not claim dwell from this data. [Primary-code] [Inference]

1.6 Transport, aggregation and missingness

  • One row per uplink. The evidence is consistent with one row per uplink; the 15-minute series is a sequence of point samples, not an average or sum over the interval.
  • Cadence is sensor-specific. Seven sensors have a median gap of 15.3–15.4 min; density-01 has 16.1 min. For the seven named sensors, 2.6–16.2% of intervals exceed 16 min; density-01 runs at a 16.1-min median, so 99.8% of its intervals exceed 16 min. Long outages reach ~2.5 h. The pooled share of >16-min gaps is 23.7%, inflated by density-01's cadence.
  • Missingness is not random. A failed LoRaWAN join, a failed TX, a battery brownout or a sensor reboot removes a row. The application sends once and sleeps; if the uplink fails, the count is not queued for a later wake. Missing rows are therefore missing observations, not zeros, and may correlate with weather, power and site conditions. [Primary-code] [Forensic]
  • LoRaWAN constraints. EU868 imposes a 1% duty cycle on the default sub-band; a 51-byte payload at SF7–SF12 has a time-on-air from tens of milliseconds to >2 s. A 15-minute uplink is well within the duty cycle, but join storms, downlinks and retransmissions can still fail. In a Brno urban measurement, the cumulative packet delivery ratio was 83%, and 50% of nodes within 1–2 km of a gateway experienced at least one packet loss. [23] [Peer-reviewed]

2. MAC address randomization

2.1 Address taxonomy

Type Top two bits Lifetime Who can resolve it Relevance to counting
Public — (OUI-assigned) Permanent Anyone; OUI identifies vendor Excluded by the CFHN LAA filter; would allow vendor/type filtering if counted
Random static 11 Until power cycle Anyone; no resolution Counted if the LAA bit is set; stable within a session
Resolvable private (RPA) 10 900 s default; 1–3600 s configurable Only holders of the IRK Counted; rotates; can be tracked via payload side channels
Non-resolvable private (NRPA) 00 Implementation-defined No one Counted; rotates; used by some Apple advertising

The locally-administered bit is set for all random types. That is why the CFHN filter counts them and excludes public addresses. [1][4]

2.2 Rotation intervals

Source Claim Confidence
Bluetooth Core Spec / SIG RPA timeout default 900 s; configurable 1–3600 s; new "Randomized RPA Updates" default range 480–900 s [Established] [2]
Apple Platform Security iOS/iPadOS use address randomization; no interval stated [Established] [4]
Android AOSP Controller/host support for RPA generation and resolution; no interval stated [Established] [5]
Becker et al. 2019 (measurement) Address lifetime: Windows 10 ~16 min; macOS/iOS ~20 min; Android 15–45 min; Fitbit Charge static (never rotates) [Peer-reviewed] [17]

Wi-Fi private-address rotation (per-network, days-to-weeks) is a separate mechanism; do not transfer its intervals to BLE. The 900 s RPA default and the 900 s wake cycle are the same order of magnitude, so a device can rotate its BLE address between consecutive bins while the person stays in place. Because the counter resets at every wake, this does not double-count within a bin; it mainly makes cross-bin and cross-sensor linkage impossible.

2.3 Effect on unique counts, dwell and re-identification

Quantity Effect of randomization Why
Within-scan unique count Little inflation Rotation interval (15–45 min) ≫ scan (10 s); most devices keep one address per scan
Across-bin count No dedup possible Bitmap resets each wake; a lingerer is counted again; rotation is irrelevant
Dwell time Not measurable No first/last-seen timestamp; rotation caps track length at ~15 min
Cross-sensor re-identification Prevented at the address level Different sensors see different addresses; no shared ID
Privacy Improved, not solved RPA can be tracked via payload carry-over (Becker: Apple 53 min, Windows 11.2 h); static-address accessories are permanent [17]

The counterintuitive result: MAC randomization is often blamed for inflating PAX counts. In this design, it does not inflate the within-scan count; the dominant overcount is repeated counting of the same lingering person across 15-minute bins, and the dominant undercount is short dwell inside a 10-s scan. Randomization mainly destroys the ability to correct either bias. This is a more precise and more damaging critique than "random MACs inflate counts." [Primary-code] [Inference]

2.4 Re-identification beyond rotation

Randomization is not anonymity. Becker, Li and Starobinski showed that BLE advertising payloads contain identifying tokens that change asynchronously with the address; an "address-carryover" algorithm tracked:

  • Windows 10 devices for up to 11.2 h across 12 address randomizations;
  • Apple devices for up to 53 min across five randomizations;
  • a Fitbit Charge indefinitely, because its address never changed.

They also found an iOS side channel that reveals user activity, and a Microsoft Surface pen that leaks an identifier. Android devices in their sample did not advertise suitable tokens, but that is an implementation detail, not a guarantee. [17] [Peer-reviewed]

For PulseCast, the implication is architectural: even if the current firmware transmits only counts, the sensor is capable of capturing raw addresses, and BLE advertising in general has been used for re-identification research. A privacy claim must therefore cover firmware integrity, update channels, physical access and retention — not only the payload.

2.5 Spoofing, cloning, static MACs and vendor OUI

  • Spoofing is trivial. An ESP32 or a phone can advertise arbitrary addresses at arbitrary intervals. A single attacker can generate hundreds of distinct 16-bit IDs in a 10-s window, inflating a sensor's count without any real footfall. There is no authentication before connection. [1] [Established]
  • Cloning is trivial. Copying a known address produces a duplicate; the 16-bit ID makes collisions easy to construct deliberately.
  • Static-address devices are invisible. The LAA filter excludes public addresses and any random-static device whose address happens to have the LAA bit clear. A Fitbit with a static random address is counted (LAA set), but a public-address phone is not. The LAA bit is not the same as the Bluetooth random-address type bits: the spec fixes the two most significant bits, not the LAA bit. If an OS does not set LAA=1 for its random addresses, the filter drops them. This is an unquantified undercount and a candidate explanation for the low raw counts. [Inference] [Uncertain]
  • Vendor OUI is unusable here. OUI is only meaningful for public addresses. Because the counter discards public addresses, it cannot distinguish a phone from a TV, a beacon, a car head-unit or a smart meter. The count is "locally-administered BLE advertisers," not "people."
  • Non-device sources. BLE is used by earbuds, watches, fitness trackers, smart home devices, beacons, cars, scooters and point-of-sale terminals. A shopping street at 19:00 contains many such devices, and their density is not proportional to pedestrian flow.

2.6 The privacy posture of the current design

The CFHN design is genuinely data-minimising:

  • Addresses are reduced to a 16-bit ID at the edge; only that ID is kept, and only for one scan.
  • Only counts leave the sensor; no raw addresses are transmitted.
  • The blog explicitly says addresses are hashed and stored for at most one hour. [6] The public 2026 code truncates rather than hashes, and keeps the ID only for one scan; the one-hour claim should be treated as unverified for the deployed firmware.

But:

  • The sensing is passive and involuntary. A bystander cannot consent, cannot see the sensor, and cannot opt out except by disabling Bluetooth.
  • The sensor is capable of raw capture; the privacy guarantee depends on firmware behaviour that a third party cannot verify from the data.
  • The 16-bit ID is not a cryptographic anonymity guarantee. It is a truncation used as a dedup key: it leaks the low 16 bits of the address, makes collisions easy to construct, and provides no cryptographic protection.
  • German data protection authorities have treated MAC-based visitor counting sceptically. The Schleswig-Holstein DPA's 2019 activity report concluded that WLAN tracking for visitor counting was unlawful; the DSK's 2022 guidance says that reading MAC addresses can fall under §25 TDDDG's consent requirement for terminal-equipment access. The legal position is contested and fact-specific, not settled. [33][34] [Gray] [Uncertain]

The honest formulation for the paper: "The current firmware minimises data, but the modality is privacy-invasive by design: it observes identifiers that people did not choose to broadcast to the operator, and its privacy properties depend on unverifiable edge behaviour." That is defensible; "the sensors are privacy-preserving" is not.


3. Failure modes

The table separates mechanisms that under-count from those that over-count and those that change the composition of the count. Directions are for the current design; some are context-dependent.

# Mechanism Direction Magnitude / evidence Confidence
1 Temporal sampling: ≈10 s scan every 900 s (1.1% duty cycle) Under-count of passers-by; over-count of lingerers across bins A person with 30 s dwell has ≈(10+30)/900 ≈ 4.4% chance of overlapping a scan; a 1 h lingerer is counted up to 4 times [Primary-code] [Inference]
2 Construct mismatch: circular presence zone, no direction Unknown; conflates flow and dwell A PAX scan counts devices present during the window, not screenline crossings; NCHRP's guidance is built around screenline counts [Established] [24]
3 Detection range / RSSI gate Under-count 20–25 m outdoors for two phones in a vineyard test near Heilbronn; RSSI threshold −80 dBm; range varies with antenna, height, enclosure [Gray] [9]
4 Body shadowing / orientation Under-count, composition bias +10 dB path loss for one body, +15 dB for two (indoor); 10 dB back vs 3 dB side at 3 m [Peer-reviewed] [20][21]
5 Crowding / collisions / queue drops Under-count at high density HCI advertising reports can be dropped when the 30-deep queue fills; dense advertising increases collisions; no drop counter in the vendored build [Primary-code] [Inference]
6 Wi-Fi coexistence Under-count BLE throughput −41% with Wi-Fi channel 6 active; PER 2–9% (up to 12%) in a shielded chamber [Gray] [22]
7 Only locally-administered addresses Under-count of public-address and LAA-clear devices; over-count of non-phone random devices Public-address phones/wearables invisible; TVs, beacons, earbuds counted; LAA bit is not the random-address type bits [Primary-code] [Inference]
8 16-bit ID collisions Under-count ≈U(U−1)/(2·65536); negligible at median, ~1.3 collisions at U=418 [Primary-code] [Inference]
9 Device multiplicity Over-count One person may carry phone + watch + earbuds; wearables can outnumber discoverable phones; exact ratio unknown for Heilbronn [Peer-reviewed] [15] [Uncertain]
10 Demographic / behavioural bias Composition bias Smartphone use 86% of 16+ in Germany but 53% of 65+; BLE representativity fell when elderly dominated; locked iPhones were not detected in the KODIS test [Official statistics] [31][32] [Gray] [9][15]
11 LoRaWAN loss / join failure Missing rows (not zeros) 2.6–16.2% of intervals >16 min for the seven named sensors; density-01 at 16.1 min; outages up to ~2.5 h; Brno urban PDR 83% [Forensic] [Peer-reviewed] [23]
12 Clock / time sync Binning / phase error Payload has no device timestamp; server receive time is used. Sleep scheduling and network latency shift the sampling phase; clock drift changes it over time [Primary-code] [Uncertain]
13 Placement / enclosure / weather Unknown, site-specific Height, orientation, glass, metal, condensation, snow; no placement metadata in the CSV [Inference]
14 Firmware / calibration change Level shift density-01 switches from integer counts to 0.7-scaled values in April 2026; mean jumps from ~5.5 to ~9.5 [Forensic]
15 Spoofing / adversarial inflation Over-count Trivial to advertise hundreds to thousands of addresses; no auth [Established] [1]
16 Cross-sensor double counting Over-count No shared ID across sensors; overlapping coverage would count the same device twice; 580 m cluster makes overlap limited but not zero [Primary-code] [Inference]

The dominant three. For a paper, do not present a flat list. Argue that three mechanisms dominate and are not fixable by calibration:

  1. Temporal sampling (1.1%) — the sensor sees a tiny, phase-random slice of each 15-minute interval. This is a design property, not noise.
  2. Construct mismatch (presence vs flow) — the count is proportional to flow × dwell, so a street with cafés and a street with commuters can produce the same count at different footfall.
  3. Unknown penetration × multiplicity — d and m move in opposite directions and are not measured. A single constant cannot track them across seasons, weather, events and demographics.

4. Evidence table

Peer-reviewed and primary sources first; gray literature is marked.

Study / source Setting Method Key finding Citation
Nicolai & Kenn (2007) Bremen (DE) and San Francisco (US) Manual gate counts + Bluetooth inquiry 2.3% (Bremen) and 6.2% (SF) of people had discoverable Bluetooth; R² 0.85–0.91 in Bremen; Bath study (O'Neill et al., as reported) ~7%, R² 0.89 [10]
Malinovskiy, Saunier & Wang (2012) Montreal and Seattle Static Bluetooth sensors + manual counts 5% (Montreal) and 2.25% (Seattle) sample rates; trends and dwell (>60 s) usable; low-volume scenarios insufficient [11]
Weppner & Lukowicz (2013) Kaiserslautern public viewing Phone-based Bluetooth scans >75% accuracy on 7 crowd-density classes using relative features; only 0.2% of people carried a scanning phone [12]
Kurkcu & Ozbay (2017) NYU transit terminal 6 Wi-Fi/Bluetooth sensors, 2 months If penetration is known, counts/flows/wait times can be accurate; reported Bluetooth detection as low as 2%; expansion depends on known penetration [13]
Angel et al. (2024) Tel Aviv, 65 detectors Calibration + speed classification 89% pedestrian detection accuracy; 60k unique MACs/day; accuracy is record classification, not volume validation [14]
Minea (2022) Bucharest subway Manual counts vs BLE/BT scans 14–49% representativity; 31.6% average at one station; age bias; wearables outnumber discoverable phones [15]
Matsuda et al. (2023) 4 indoor spaces, Japan BLE scanners + camera/manual labels MAE 1.4–5.5 people; MAPE 24–91%; performance poor below 10 people [16]
Becker, Li & Starobinski (2019) Lab/field BLE capture Address + payload tracking Address life ~16 min (Windows), ~20 min (Apple), 15–45 min (Android); carry-over tracking 53 min (Apple), 11.2 h (Windows); Fitbit static [17]
Martin et al. (2017) 2.6M Wi-Fi MAC corpus Wi-Fi randomization study iOS randomized; Android adoption low in 2015–16; 96% of Android defeatable [18]
Fenske et al. (2021) 160 phone models Lab Wi-Fi randomization test Newer phones are better; many older phones still leak; randomization is device/OS-dependent [19]
Januszkiewicz (2018) Indoor 2.4 GHz FDTD + measurements +10 dB path loss for one body, +15 dB for two; 10–20 dB near body [20]
Deng, Fu & Wang (2018) Indoor BLE RSSI measurements 10 dB back-shadowing, 3 dB side at 3 m; 6–8 dB typical BSIE [21]
Mahmud et al. (2023), INL Anechoic chamber BLE + Wi-Fi coexistence BLE throughput −41% under Wi-Fi; PER 2–9% (up to 12%) [22]
Masek et al. (2021) Brno, 311 locations LoRaWAN urban measurement 83% cumulative PDR; 50% of nodes within 1–2 km had packet loss [23]
Ryus et al. (2014), NCHRP 797 US Standards/methods review Correction factors from validation counts; annual revalidation; accuracy vs consistency; screenline vs presence [24]
FHWA TMG Ch. 4 US Federal guidance Manual counts as ground truth; automated counters need calibration; short counts are highly variable [25]
Ohlms et al. (2018), VTRC 19-R4 Virginia State program review Bluetooth alone not feasible as of 2016; reported Wi-Fi detection ~26.4%; validation produces site-specific correction factors [26]
Hook & Regal (1995, 2000) Epidemiology Capture-recapture theory Four assumptions; positive dependence biases estimates downward; empirical underestimation 16–25% [27][28]
Code for Heilbronn (2025–26) Heilbronn Blog + firmware repo Heltec ESP32 + SX1262; BLE-only; LAA filter; 16-bit ID (truncation); 0.7 factor; 900 s sleep; one uplink per wake [6][7]
Fraunhofer IAO/KODIS (2023) Heilbronn vineyards 3 LoPy4 + 2 phones BLE range 20–25 m; locked iPhone not detected; simultaneous Wi-Fi+BLE causes errors; trend use only [9]
Destatis (2024); Bitkom (2026) Germany Household/person surveys 90.8% of households have a smartphone; 86% of 16+ use one; 53% of 65+ [31][32]

What the table establishes. Across four decades and three radio generations (classic Bluetooth inquiry, Wi-Fi probe requests, BLE advertising), the same pattern recurs: detection is a small, biased, variable sample of people; counts track trends; absolute levels require external calibration. The Heilbronn deployment is not an outlier; it is a particularly sparse instance of a known design.

What the table does not establish. No study validates a 10-s BLE scan every 15 min against manual counts in a German shopping street. No study measures the device-multiplicity distribution in Heilbronn. No study quantifies the effect of the 0.7 factor. These are gaps, not evidence of accuracy.


5. Calibration and validation

5.1 What counts as ground truth?

Method Strength Weakness Use for BLE
Manual screenline counts Direct, transparent, cheap Labour-intensive; 2 h fatigue; observer bias; no direction at high flow Best short-term reference; stratify by daypart and day type
Video with manual annotation Auditable; can rewind; direction Privacy; storage; occlusion in crowds Best for dense sites; annotate a sample, not 24/7
Passive infrared / laser beam Continuous; cheap Occlusion; group counting; no identity Independent continuous reference, but has its own bias
Turnstiles / gate counts Exact at controlled entries Only for enclosed sites Good for a shop entrance, not a street
Wi-Fi probe counts Higher detection than BLE in some studies Randomization; OS differences; also privacy-invasive Not ground truth; a second biased sensor

The NCHRP guidance is explicit: accuracy is the magnitude of difference from a ground-truth count; consistency is the residual variability after correcting expected bias; correction factors should be developed locally and revalidated at least annually. [24] The FHWA TMG likewise treats manual counts as the reference and warns that short non-motorized counts are highly variable. [25]

5.2 Ratio / expansion estimation

The simplest calibration is a ratio estimator:

P_hat = y / r,     r = mean(y / P) over validation samples

Equivalently, P_hat = k · y with k = 1/r. This is the "penetration rate" or "expansion factor" approach used in Bluetooth transport studies; the ratio estimator itself is standard sampling theory. [11][13][30]

What it requires: a stable r across the conditions where it is applied. What breaks it:

  • r varies with device mix (iOS/Android/wearables), age, weather, time of day, event type and sensor placement.
  • r is a ratio of two noisy quantities; its variance is dominated by small P bins. At P = 10, a ±3 count error is ±30%.
  • A single global k applied to all sensors hides site-specific detection zones.
  • If r is estimated from a period with different firmware/factor, it is not transferable.

NCHRP's guidance is explicit that a correction factor is only valid for the site and conditions where it was developed; local factors and auxiliary counts are preferred over a single global constant. [24] The same logic applies to BLE: a global 0.7 cannot absorb site, season and device-mix variation. The lesson is not "Bluetooth is useless"; it is "a single expansion factor is not enough; you need auxiliary counts and a model."

5.3 Capture-recapture

With two or more independent observations of the same population, the Lincoln–Petersen estimator is:

N_hat = (n1 · n2) / m

where n1, n2 are the counts in each sample and m is the number matched. Four assumptions must hold: closed population, correct matching, independence of samples, and homogeneous capture probability. [27]

Why this is hard for BLE:

  • Matching requires a persistent identifier. The current firmware has none; even with persistent IDs, RPA rotation breaks matching after ~15 min.
  • Independence is violated if both sensors see the same bodies/obstacles or if a person carrying more devices is more likely to be captured by both.
  • Heterogeneity is severe: a person with a phone + watch has multiple chances; a person with a locked iPhone may have none. Heterogeneous capture probabilities can bias N_hat downward; positive dependence also biases it downward. Hook & Regal found 16–25% underestimation in empirical capture-recapture studies with positive dependence. [28]

Verdict: capture-recapture is a useful diagnostic (e.g., two adjacent sensors with overlapping coverage, or manual + sensor), but naive Lincoln–Petersen on BLE counts will not give a defensible population estimate. If used, report the assumptions and a sensitivity range, not a point estimate.

5.4 Regression calibration

Regress ground truth on sensor counts:

P_t = α + β · y_t + γ · X_t + ε_t

where X_t includes hour, day type, weather, sensor ID and event flags. Use a time-based holdout and report out-of-sample MAE/MAPE, not in-sample R².

Why it can help: it allows the detection rate to vary with covariates; it can absorb site-specific factors; it gives a principled way to propagate uncertainty.

Why it can mislead:

  • Classical measurement error in y attenuates β toward zero; the naive model looks less responsive than it is. Regression calibration replaces y with E[P|y] and can correct this, but only if the calibration model is correctly specified and validated on independent data. [29]
  • If the sensor's detection rate changes after the calibration period (firmware, device mix, season), the calibration is stale.
  • Extrapolating beyond the observed range (e.g., event peaks) is unsupported.
  • A high R² on a holdout does not prove the level is right; it proves the variation is tracked.

For PulseCast, the honest use is: calibrate the index to people for a defined site and period, report the calibration window and covariates, and revalidate. Do not apply a June calibration to November without evidence.

5.5 What can and cannot be inferred

Question Can the current data answer it? Why
Is today busier than yesterday? Yes, as a relative index Same sensor, same unit, short horizon; detection conditions similar
What is the hourly/daily profile? Yes, descriptively The series has a coherent daily/weekly shape; high lag-1 autocorrelation (0.84–0.95 at 15 min)
How many people passed in the last 15 min? No y = 0.7 × U, where U is a 10-s snapshot count; not an integrated flow
How many people are in the area now? No, not without calibration U is devices, not people; d, m, q unknown
How long do people stay? No No first/last-seen; dedup resets; RPA rotation
Which sensor has more footfall? Only with equal detection zones and factors Placement, range and factor differ; the 0.7 factor is global but the environment is not
Did an event increase footfall? Suggestive Relative change is visible; magnitude is not identified
Are older people under-represented? Yes, likely 53% smartphone use among 65+ vs 86% overall; BLE representativity falls with elderly share [15][31][32]
Can we forecast next week? Yes, in device units The series is predictable; the forecast inherits the unit error
Can we claim privacy-preservation? Only with architecture + governance evidence Payload is minimised; sensing is involuntary; firmware is unverifiable

5.6 A minimum validation protocol for PulseCast

If the paper wants to make a calibration claim, this is the smallest defensible protocol:

  1. Sites: at least 3 sensors spanning the range of environments (shopping street, riverfront, mall entrance).
  2. Reference: manual screenline counts by two observers, 15-min bins, 2 h per session; video for dense bins.
  3. Design: 2 weeks; 4 dayparts (morning, lunch, afternoon, evening); 2 weekdays + 1 weekend day per week; at least 200 people per bin for stable ratios.
  4. Synchronisation: align sensor timestamps to the manual bin; discard bins with a gap > 1 min.
  5. Metrics: ratio y/P, bias, MAE, MAPE, Bland–Altman limits of agreement; report by site and daypart, not only pooled.
  6. Model: calibration with hour, day type, weather and sensor ID; time-based holdout; report out-of-sample error.
  7. Uncertainty: bootstrap the ratio; propagate the calibration interval into the staffing recommendation.
  8. Revalidation: repeat at least seasonally; re-check after firmware or device-mix changes.
  9. Report: publish the raw U, the factor, the calibration window and the residual diagnostics. Do not publish only the calibrated people number.

This is more work than a hackathon usually allows, but even a one-day, three-site pilot would be enough to replace "we assume 0.7" with "we measured a ratio of X (95% CI Y) on one day in November." That is a much stronger paper.


6. Implications for PulseCast

6.1 Reframe the target

The current target is a 15-minute "density" value. The defensible target is a relative activity index: the number of distinct locally-administered BLE advertising addresses seen in a short scan, scaled by a documented constant. Forecast that index, and separately report a calibrated people estimate only where validation exists.

Suggested wording for the paper:

"The sensor reports y = 0.7 · U, where U is the number of distinct locally-administered BLE advertising addresses observed during a single ~10-second scan. U is a device-address count, not a person count. We treat y as a relative activity index and evaluate forecasts in index units. A people-level interpretation would require external calibration, which this dataset does not contain."

This is honest, precise and still supports the forecasting contribution.

6.2 Keep the forecasting, change the claim

The series is predictable: daily and weekly cycles, weather sensitivity, holiday effects. A global gradient-boosted model with time-based holdout is a reasonable method. The forecast skill is real in index units. The current model's holdout metrics (MAE 3.36, WAPE 27.8%, R² 0.779; 7-day MAE 4.37, WAPE 36.2%, R² 0.672) are computed on y, so they measure index forecast skill; a people-level error cannot be derived from them without a calibration study. The paper can claim:

  • "We forecast the sensor activity index with MAE X, WAPE Y%, R² Z, against a seasonal-naive baseline."
  • "The weather ablation shows temperature/rain add information beyond the time features."
  • "The staffing heuristic translates index units into staff hours under explicit assumptions."

It cannot claim:

  • "We forecast footfall."
  • "The model estimates visitors."
  • "The staffing plan is calibrated to people."

unless a validation study is added.

6.3 Sensitivity analysis for the 0.7 factor

Because f = 0.7 is a default, not a measurement, the paper should show how conclusions change under plausible alternatives. A simple approach:

  • Recompute the staffing plan for f ∈ {0.4, 0.7, 1.0, 1.4} (or a range implied by the literature).
  • Report the range of staff hours, not a single number.
  • State that the ranking of hours is invariant to a global multiplicative factor, but the level and the cost are not.

This turns an unidentifiable constant into a transparent assumption. It also protects against the reviewer objection "why 0.7?"

6.4 Do not compare levels across sensors

The eight sensors differ in placement, range, enclosure, cadence (density-01 at 16.1 min vs 15.3 min), and possibly factor. A ranking of "busiest sensor" is only valid if detection zones and factors are equal. The paper can compare temporal shapes (when each site peaks) more defensibly than levels (how many people each site has).

If a heat map is required, label it "activity index, not footfall" and avoid summing across sensors (sums double-count if coverage overlaps and mix units).

6.5 What a privacy-first alternative must demonstrate

The paper argues for a privacy-first alternative. To be credible, the alternative should be evaluated on the same axes as the PAX counter:

Axis PAX counter (current) Requirement for alternative
Unit Device-address count × 0.7 People or clearly defined physical quantity
Sampling 10 s / 900 s (1.1%) Continuous or documented sampling; no phase bias
Direction None Direction or screenline crossing
Dwell Not measurable Measurable or explicitly out of scope
Identity 16-bit address ID at edge No device identifiers at all
Consent Impossible Notice + opt-out, or genuinely anonymous sensing
Accuracy Unvalidated Validated against manual/video ground truth
Failure modes Range, body, crowding, weather, LoRaWAN Documented and quantified
Cost Low Comparable total cost (hardware + install + maintenance)
Governance Firmware-dependent Auditable, open, DPIA-compatible

Candidate modalities that can meet these requirements better: thermal array counting, radar/LiDAR, stereo vision with on-device counting (no images stored), passive infrared beams at entrances, or manual/video sampling with a statistically designed schedule. Each has its own failure modes; the paper should compare, not assert.

6.6 What the paper should say — and not say

Say:

  • "The deployed sensors count locally-administered BLE advertising addresses in short scans; the published value is 0.7 × that count."
  • "This is a relative index; absolute people counts are not identified without external calibration."
  • "The 15-minute series is a sequence of point samples, not an aggregate."
  • "MAC randomization and the absence of cross-scan dedup prevent dwell estimation and cross-sensor deduplication."
  • "The privacy claim depends on edge firmware behaviour and on a legal basis that is contested."
  • "Our forecast is evaluated in index units; the staffing plan is a decision aid under explicit assumptions."

Do not say:

  • "The sensors count people."
  • "The data is 15-minute footfall."
  • "0.7 converts devices to people." (It is a default, not a validated conversion.)
  • "The sensors are privacy-preserving." (Say "data-minimising" and explain the residual risks.)
  • "MAC randomization inflates the counts." (In this design, the dominant bias is temporal sampling and repeated counting of lingerers, not within-scan rotation.)
  • "The model generalises to other cities." (It is a single 580 m cluster, 5 months, 8 sensors.)

7. Open questions

  1. Which firmware ran in 2025? The 2025 blog describes a one-hour hashed store; the 2026 code resets every scan. Resolve by obtaining the deployed binary or the git history/tags for sensor-pax. Why it matters: the 2025–Mar 2026 density-01 segment may have a different dedup window and factor.
  2. What are the deployed factors? The default is 0.7; provisioning requires a value. Are all sensors 0.7? Were any calibrated? Why it matters: a per-sensor factor changes cross-sensor comparisons.
  3. What is the actual scan duration in the field? The code suggests ≈10 s; the configured value is 30 s. Measure with a BLE advertiser of known rate plus a current/logic trace, or instrument the firmware. Why it matters: the duty cycle (1.1% vs 3.3%) changes the undercount factor by 3×.
  4. What is the device-multiplicity distribution in Heilbronn? Phone + watch + earbuds vs phone only; locals vs tourists; age groups. Why it matters: m is the main overcount; d is the main undercount; their product is unknown.
  5. What is the share of locked/standby phones that do not advertise? The KODIS test found a locked iPhone SE was not detected; is that reproducible across models? Why it matters: it sets the effective d.
  6. How much do sensors overlap? With 20–25 m range and a 580 m cluster, overlap is likely limited, but adjacent sensors may see the same device. Why it matters: cross-sensor double counting.
  7. Is missingness random? Compare missing intervals with weather, battery voltage and gateway logs. Why it matters: if outages correlate with heat/storm, the series is biased.
  8. What is the legal basis in Baden-Württemberg? A DPIA, a documented legitimate-interest assessment, and signage/notice. Why it matters: the paper's privacy argument needs a jurisdiction-specific anchor, not a generic GDPR gesture.
  9. Can a privacy-first sensor match the cost? Thermal/radar/video vs ESP32 + LoRaWAN. Why it matters: the alternative must be deployable at 8+ sites, not only in a lab.
  10. What would falsify the "relative index" claim? A manual validation showing that the index's ratio to people varies systematically by daypart or weather. Why it matters: it defines the paper's own success criterion.

References

[1] Bluetooth SIG. Bluetooth Core Specification v5.3. 2021. https://www.bluetooth.com/specifications/specs/core-specification-5-3/ — advertising channels, intervals, address types, scan parameters.

[2] Bluetooth SIG. "Enhancing device privacy and energy efficiency with Bluetooth Randomized RPA Updates." 15 May 2025. https://www.bluetooth.com/blog/enhancing-device-privacy-and-energy-efficiency-with-bluetooth-randomized-rpa-updates/ — RPA timeout default 900 s, range 1–3600 s; randomized update range 480–900 s.

[3] Bluetooth SIG. Understanding Reliability in Bluetooth Technology. 2020. https://www.bluetooth.com/wp-content/uploads/2020/10/EN-Understanding_Reliability.pdf — three primary advertising channels; random 0–10 ms delay; collision avoidance.

[4] Apple. "Bluetooth security." Apple Platform Security. https://support.apple.com/guide/security/bluetooth-security-sec82597d97e/web — BLE address randomization; no interval stated.

[5] Android Open Source Project. "HCI requirements." https://source.android.com/docs/core/connect/bluetooth/hci_requirements — RPA generation/resolution support; address type fields.

[6] Code for Heilbronn. "Passantenzahlen messen mit LORA." 18 September 2025. https://codeforheilbronn.de/blog/2025/09/passantenzahlen-messen-mit-lora/ — hardware, claimed hashing, one-hour retention claim, LoRaWAN.

[7] Code for Heilbronn. lorax / lora-frequenzmessung, branch develop. https://codeberg.org/cfhn/lorax — firmware: sensor-pax/src/main.cpp, pax.h, provision.h; vendored sensor-pax/lib/paxlib/libpax.cpp, blescan.cpp, libpax_api.cpp. Accessed 14 November 2026.

[8] Fraunhofer IAO / KODIS. "Frequenzmessung im innerstädtischen Bereich." https://www.kodis.iao.fraunhofer.de/de/projekte/frequenzmessung.html — ESP32 Paxcounter deployment in Heilbronn and Schwäbisch Hall.

[9] Straßburg, S., Buhl, D., Ködel, L., Göbels, V.P., & Schaefer, K. "Passenger Frequency Measurement — Part I." Research Center KODIS, 11 December 2023. https://medium.com/research-center-kodis/passenger-frequency-measurement-part-i-6a75d2aa63fd — 20–25 m BLE range; locked iPhone not detected; Wi-Fi/BLE interference; trend use only.

[10] Nicolai, T., & Kenn, H. (2007). "About the relationship between people and discoverable Bluetooth devices in urban environments." Mobility '07, 72–78. DOI: 10.1145/1378063.1378076.

[11] Malinovskiy, Y., Saunier, N., & Wang, Y. (2012). "Analysis of pedestrian travel with static Bluetooth sensors." Transportation Research Record 2299:137–149. DOI: 10.3141/2299-15.

[12] Weppner, J., & Lukowicz, P. (2013). "Bluetooth based collaborative crowd density estimation with mobile phones." IEEE PerCom 2013, 193–200. DOI: 10.1109/PerCom.2013.6526732.

[13] 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. DOI: 10.3141/2644-09.

[14] Angel, A., Cohen, A., Dalyot, S., & Plaut, P. (2024). "Estimating pedestrian traffic with Bluetooth sensor technology." Geo-spatial Information Science 27(5):1391–1404. DOI: 10.1080/10095020.2023.2247446.

[15] Minea, M. (2022). "An experimental assessment of people's location efficiency using low-energy communications-based movement tracking." Sensors 22(22):9025. DOI: 10.3390/s22229025.

[16] Matsuda, Y., Ueda, K., Taya, E., Suwa, H., & Yasumoto, K. (2023). "BLECE: BLE-based crowdedness estimation method for restaurants and public facilities." ICMU 2023, 1–6. DOI: 10.23919/ICMU58504.2023.10412158.

[17] Becker, J.K., Li, D., & Starobinski, D. (2019). "Tracking anonymized Bluetooth devices." Proceedings on Privacy Enhancing Technologies 2019(3):50–65. DOI: 10.2478/popets-2019-0036.

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

[19] Fenske, E., Brown, D., Martin, J., Mayberry, T., Ryan, P., & Rye, E. (2021). "Three years later: A study of MAC address randomization in mobile devices and when it succeeds." Proceedings on Privacy Enhancing Technologies 2021(3):164–181. DOI: 10.2478/popets-2021-0042.

[20] Januszkiewicz, Ł. (2018). "Analysis of human body shadowing effect on wireless sensor networks operating in the 2.4 GHz band." Sensors 18(10):3412. DOI: 10.3390/s18103412.

[21] Deng, Z., Fu, X., & Wang, H. (2018). "An IMU-aided body-shadowing error compensation method for indoor Bluetooth positioning." Sensors 18(1):304. DOI: 10.3390/s18010304.

[22] Mahmud, S.A., Kasera, S., Ji, M., & Agarwal, V. (2023). Experimental Evaluation of Interference in 2.4 GHz Wireless Network. INL/RPT-23-74719, Idaho National Laboratory. https://www.osti.gov/servlets/purl/2242485.

[23] Masek, P., Stusek, M., Svertoka, E., Pospisil, J., Burget, R., Lohan, E.S., Marghescu, I., Hosek, J., & Ometov, A. (2021). "Measurements of LoRaWAN technology in urban scenarios: A data descriptor." Data 6(6):62. DOI: 10.3390/data6060062.

[24] Ryus, P., Ferguson, E., Laustsen, K.M., Schneider, R.J., Proulx, F.R., Hull, T., & Miranda-Moreno, L. (2014). Guidebook on Pedestrian and Bicycle Volume Data Collection. NCHRP Report 797, Transportation Research Board. https://escholarship.org/content/qt11q5p33w/qt11q5p33w.pdf.

[25] Federal Highway Administration. Traffic Monitoring Guide, Chapter 4: "Traffic Monitoring for Non-Motorized Traffic." 2013. https://www.fhwa.dot.gov/policyinformation/tmguide/tmg_2013/traffic-monitoring-for-non-motorized.cfm.

[26] Ohlms, P.B., Dougald, L.E., & MacKnight, H.E. (2018). Assessing the Feasibility of a Pedestrian and Bicycle Count Program in Virginia. VTRC 19-R4, Virginia Transportation Research Council. https://vtrc.virginia.gov/media/vtrc/vtrc-pdf/vtrc-pdf/19-R4.pdf.

[27] Hook, E.B., & Regal, R.R. (1995). "Capture-recapture methods in epidemiology: Methods and limitations." Epidemiologic Reviews 17(2):243–264. DOI: 10.1093/oxfordjournals.epirev.a036192.

[28] Hook, E.B., & Regal, R.R. (2000). "Accuracy of alternative approaches to capture-recapture estimates of disease frequency: Internal validity analysis of data from five sources." American Journal of Epidemiology 152(8):771–778. DOI: 10.1093/aje/152.8.771.

[29] Carroll, R.J., Ruppert, D., Stefanski, L.A., & Crainiceanu, C.M. (2006). Measurement Error in Nonlinear Models: A Modern Perspective, 2nd ed. Chapman & Hall/CRC. ISBN 978-1-58488-633-4.

[30] Cochran, W.G. (1977). Sampling Techniques, 3rd ed. Wiley. ISBN 978-0-471-16240-7.

[31] Statistisches Bundesamt (Destatis). "Ausstattung privater Haushalte mit Informations- und Kommunikationstechnik." 2024. https://www.destatis.de/DE/Themen/Gesellschaft-Umwelt/Einkommen-Konsum-Lebensbedingungen/Ausstattung-Gebrauchsgueter/Tabellen/liste-gebietsstaende-Infotechnik.html.

[32] Bitkom Research. "Smartphone-Markt 2026." 2026. https://bitkom-research.de/studien/smartphone-markt-2026 — 86% of 16+ use a smartphone; 53% of 65+.

[33] Unabhängiges Landeszentrum für Datenschutz Schleswig-Holstein. 37. Tätigkeitsbericht 2019, pp. 97–99. https://www.datenschutzzentrum.de/tb/tb37/uld-37-taetigkeitsbericht-2019.pdf — WLAN tracking for visitor counting considered unlawful.

[34] Datenschutzkonferenz (DSK). Orientierungshilfe für Anbieter digitaler Dienste. 2022, para. 22. https://www.datenschutzkonferenz-online.de/media/oh/OH_Digitale_Dienste.pdf — reading MAC addresses can require consent under §25 TDDDG.

[35] Yung, C. (2024). A machine learning framework for public transport ridership estimation using multi-source data fusion with low-cost Bluetooth data. Master's thesis, Aalto University. https://urn.fi/URN:NBN:fi:aalto-202501241673 — BLE-only model R² = −0.49; fused model R² = 0.86. [Gray/thesis]


Appendix A — Reproduction

The forensic claims are reproducible from the delivered CSV with standard Python. No external data are needed.

import csv, datetime as dt, statistics, collections

rows = [r for r in csv.DictReader(open("data/raw/schwarz-pax.csv"))
        if r["type"] == "densityData"]

start = dt.datetime.fromisoformat("2026-06-12T00:00:00+00:00")
end   = dt.datetime.fromisoformat("2026-11-08T00:00:00+00:00")
prod  = [r for r in rows
         if start <= dt.datetime.fromisoformat(r["time"]) < end
         and r["name"] not in ("maker-fair", "niklas-01", "hn-bollwerksturm-01")]

vals = [float(r["value"]) for r in prod]
raw  = [round(v / 0.7) for v in vals]

# 1. All production values are exact multiples of 0.7
assert all(abs(v / 0.7 - round(v / 0.7)) < 1e-6 for v in vals)

# 2. Reconstructed integer counts
print("rows:", len(vals))
print("median U:", statistics.median(raw))
print("p95 U:", sorted(raw)[int(0.95 * len(raw))])
print("max U:", max(raw))
print("zero share:", raw.count(0) / len(raw))

# 3. Per-sensor multiples-of-0.7 check (production window only)
by = collections.defaultdict(list)
for r in prod:
    by[r["name"]].append(float(r["value"]))
for name, v in by.items():
    share = sum(abs(x / 0.7 - round(x / 0.7)) < 1e-6 for x in v) / len(v)
    print(f"{name:26s} n={len(v):6d} multiples0.7={share:.1%}")

# 4. The old density-01 hardware is integer-valued, not 0.7-scaled
old = [r for r in rows if r["name"] == "density-01"
       and dt.datetime.fromisoformat(r["time"]) < dt.datetime.fromisoformat("2026-04-01T00:00:00+00:00")]
old_vals = [float(r["value"]) for r in old]
print("old density-01 integer share:",
      sum(abs(v - round(v)) < 1e-6 for v in old_vals) / len(old_vals))

Expected output for the production window: rows: 83156, median U: 12, p95 U: 71, max U: 418, zero share: ~0.027, and multiples0.7=100.0% for all eight production sensors. The older density-01 hardware (863f75b0, 2025-10 → 2026-03) is 100% integer-valued; the replacement (f8c19730, 2026-04 → 2026-11) is 100% multiples of 0.7.


Appendix B — Source map for the strongest claims

Claim Source
10-s scan every 900 s; 16-bit ID (truncation); LAA filter; 0.7 factor sensor-pax/src/pax.h, main.cpp, provision.h; lib/paxlib/libpax.cpp, blescan.cpp, libpax_api.cpp [7]
100% of production values are multiples of 0.7 Forensic, Appendix A
density-01 integer-to-0.7 regime break Forensic, Appendix A; monthly means in Section 1.4
RPA default 900 s, range 1–3600 s Bluetooth SIG 2025 [2]
Address lifetimes 15–45 min Becker et al. 2019 [17]
20–25 m range; locked iPhone not detected KODIS blog [9]
10–20 dB body loss Januszkiewicz 2018 [20]; Deng et al. 2018 [21]
BLE throughput −41% under Wi-Fi INL report [22]
2–7% classic BT discoverability Nicolai & Kenn 2007 [10]; Malinovskiy et al. 2012 [11]
14–49% BLE representativity; age bias Minea 2022 [15]
MAE 1.4–5.5 people; MAPE 24–91% Matsuda et al. 2023 [16]
BLE-only R² = −0.49 Yung 2024 [35]
83% LoRaWAN PDR in Brno Masek et al. 2021 [23]
Capture-recapture assumptions and downward bias Hook & Regal 1995/2000 [27][28]
NCHRP correction-factor methodology Ryus et al. 2014 [24]
German legal uncertainty ULD 2019 [33]; DSK 2022 [34]