Real-Time AI-Driven People Tracking and Counting Using Overhead Cameras

Benny Kai Kiat Ng; Chamith Dilshan Ranathunga; Chau Yuen; Dinuka Sandun Udayantha; Ishrath Ahamed

REVIEW 5 major objections 4 minor 12 references

Real-Time AI-Driven People Tracking and Counting Using Overhead Cameras

T0 review · 5 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A single overhead camera can count people entering and exiting a doorway in real time at 97% accuracy while running on a low-power edge computer.

desk verdict A sensible edge-counting system whose headline 97% accuracy is contradicted by its own 92% long-term test and is not reproducible from the text. read the letter →

arxiv 2411.10072 v1 pith:PVNOOKA6 submitted 2024-11-15 cs.CV cs.AI

Ishrath Ahamed , Chamith Dilshan Ranathunga , Dinuka Sandun Udayantha , Benny Kai Kiat Ng , Chau Yuen This is my paper · ORCID

classification cs.CVcs.AI

keywords peoplecountingoverheadcameraobjecttrackingedgecomputingheaddetectionSSDMobileNetfeaturematchingregionhistory

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to show that a single overhead camera and a low-power edge computer can count people entering and exiting a doorway in real time without sacrificing accuracy for speed. The proposed pipeline combines a fine-tuned SSD MobileNet head detector, a MobileNetV2 feature extractor that produces a 1024-dimensional embedding per head, a custom feature-distance tracking algorithm, and a three-zone counting rule where movement from the outside zone to the inside zone registers an entry and the reverse registers an exit. The authors report 97% counting accuracy at 20-27 FPS on an Intel NUC mini PC with no GPU, which would beat the accuracy-FPS trade-off of prior door-monitoring systems. A two-day long-term test reports an overall accuracy of 92%, with per-day values of 89.66% and 95.24%, so the headline figure refers to separate live video testing.

What carries the argument

The load-bearing component is the region-history attribute attached to each tracked head object, together with the assignment-based tracking algorithm. The camera's field of view is divided into three horizontal zones: A (outside), B (critical boundary), and C (inside); each person-object stores the sequence of zones it has occupied, and an entry is counted only when the sequence moves from A to C, an exit only from C to A. Tracking is framed as an assignment problem on an $m \times n$ feature-distance matrix $M$, where the algorithm greedily pairs the closest registered and new detections subject to a distance threshold $T$ and a spatial cap $D$, and each unassigned old object increments a miss counter $eCount$ that removes the object after $E$ consecutive misses. The head detector is a fine-tuned SSD MobileNet with separate day and night models, and the feature extractor is a compressed MobileNetV2 that takes a $120\times120\times3$ crop and outputs a $1\times1024$ embedding.

What would settle it

Take a known ground-truth scenario where people step into the doorway's boundary zone and then step back out without fully entering, and compare the system's reported count with the true number of entries and exits; if the system registers changes that did not happen, the counting rule fails exactly in the lingering-and-reversing case.

Watch

Extended reading notes

Core claim

The central claim is that the combination of a fine-tuned single-shot head detector, a lightweight embedding-based tracker, and a three-region history rule solves the accuracy-versus-speed problem in people counting. On the paper's terms, the method is state-of-the-art because it keeps accuracy at 97% while running at 20-27 FPS on edge hardware, where prior work either ran fast with lower accuracy (86.32% at 45 FPS) or achieved high accuracy at impractically low frame rates (93.10% at 0.6 FPS, 95% at 5 FPS). The counting rule is designed to avoid the oscillation errors of two-zone systems by requiring a full outside-to-inside transition before registering an entry, and the tracking algorithm tolerates short detection gaps through a miss counter. The long-term laboratory test shows the system maintains a 92% overall accuracy across two days of continuous operation.

Load-bearing premise

The 97% figure rests on the assumptions that people cross the doorway in one direction without reversing in the boundary zone and that the detector never drops a person long enough to hit the miss threshold E.

Editorial extensions

If this is right

At 20-27 FPS on a GPU-less mini PC, the system can be deployed for continuous door monitoring in office buildings, retail stores, and transit stations where power and cost are constrained.
The day/night dual-model detection means the same hardware can provide 24-hour counting without additional lighting, using the camera's IR mode to switch models.
Because tracking is feature-based rather than bounding-box-based, the method tolerates occlusions and brief detection misses as long as they stay under the E-frame threshold.
In long-term operation, the method reports 92% overall accuracy, with the two-day test showing errors can be as low as one miscount per day when traffic is light.
The accuracy-FPS combination implies that prior trade-offs, such as 93% at 0.6 FPS or 95% at 5 FPS, are not fundamental limits for overhead-camera counting.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural stress test is to have people linger in the boundary zone and reverse direction; the paper does not specify how the region history handles A-B-A paths, so this is the most likely place for the 97% claim to break down in practice.
The reported 97% live accuracy versus 92% long-term accuracy suggests the figure is scenario-dependent; reporting error breakdowns by entry vs exit and by crowd size would make the method's true operating envelope clear.
The same assignment-plus-region-history logic could be transferred to other top-down counting tasks, such as objects on conveyor belts or vehicles in lanes, by retraining the detector on the new object class.
The dual day/night model selection could be replaced by an on-the-fly brightness estimator, making the system camera-agnostic rather than tied to the IR-mode signal of one camera model.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Thanks for the report. My take: this is a coherent, modest engineering paper that overstates its one headline number. The 97% accuracy in the abstract is contradicted by the paper's own Table II, which reports 92% overall across two days (4 errors on 50 observations). The live test that supposedly hit 97% has no described protocol—no event count, no error definition, no scene description—so the number is not reproducible from the manuscript.

What the paper does well: it builds a sensible low-cost overhead-camera people counter using SSD-MobileNet head detection, MobileNetV2 embeddings, and a greedy feature-distance tracker with a miss-count threshold. The practical touches are real: fine-tuning the detector to reject chairs and bags, using the camera's IR mode to switch between day and night detection models, and downsampling inputs for edge inference. The FPS figures (20-27 on a NUC CPU) are believable for this pipeline. The two-day long-term monitoring is the most credible part of the evaluation, but it shows 92%, not 97%.

The soft spots are where the claims live. The counting logic in Section II-D is underspecified at exactly the tricky part. The rule says an entry is A-to-C and an exit C-to-A, with B as a 'critical' region, but there is no state machine for A->B->A, A->B->C->B->C, starting inside B, or detection dropout splitting a crossing into two IDs. Algorithm 1's thresholds T, D, and E are never given. The SOTA comparison is three entries, one of which is the same group's prior work, so the '2% improvement' is not independently established. No modern trackers (DeepSORT, ByteTrack) are compared, and no public benchmark is used.

If you weigh the central claim, the paper fails its own evidence bar. But the engineering is mostly sensible, and the two-day data gives a starting point. A serious referee could ask for the missing protocol, threshold values, and trajectory handling, and the paper might become useful after heavy revision. I would not cite it as is, but I would not mind seeing a revised version. For peer review: yes, send it out—not because it is ready, but because a competent reviewer can turn the vague accuracy claim into a concrete list of missing details. For a reading group, maybe, as an example of how system papers should report test protocols.

Referee Report

5 major / 4 minor

Summary. The paper proposes a real-time people-counting system based on an overhead camera. The pipeline consists of a fine-tuned SSD MobileNet head detector (day and night models), a MobileNetV2-based feature extractor, a custom feature-matching tracker (Algorithm 1), and a three-region counting rule (Section II-D). The authors report a 97% counting accuracy and 20–27 FPS on an Intel NUC edge computer, and compare favorably with prior works in Table III.

Significance. If the claimed 97% accuracy at 20–27 FPS on a low-power edge device were substantiated, the system would be a practically useful contribution to smart-building and safety applications, especially because the method aims to handle crowded scenes and low-light conditions. The paper, however, provides no reproducible experimental protocol, no dataset release, and no code release, and its own Table II reports 92% accuracy in the only described controlled test. The counting algorithm is underspecified at the exact point where real door-crossing trajectories are nontrivial. These deficiencies make the central performance claim unverifiable from the manuscript as written.

major comments (5)

[Section II-D] The counting rule is underspecified. The text defines an entry as movement from Region A to Region C and an exit from C to A, with B as a 'critical' intermediate region, but it never defines how partial or oscillating paths are handled. In particular, the manuscript does not state whether A→B→A is ignored, whether A→B→C→B→C yields one count or two, how an object first detected in B is treated, or what debounce/hysteresis rule prevents repeated counting at the boundary. Since the counting accuracy is the central claim of the paper, this missing state machine makes the reported 97% accuracy irreproducible.
[Section III, Table II] The abstract and Section III claim an overall accuracy of 97% from 'live video testing,' yet the only controlled test described in detail, the two-day continuous run in Table II, reports an overall accuracy of 92.00% (4 errors on 50 actual observations, with predicted totals 26/26 versus actual 26/24). No event counts, duration, ground-truth protocol, or error breakdown are given for the live test that allegedly achieved 97%. The discrepancy between 97% and 92% is unexplained and must be resolved with a full description of both test protocols.
[Section II-C, Algorithm 1] The tracking algorithm depends on several free parameters whose values and selection procedure are never reported: the feature-distance threshold T, the spatial-distance threshold D, and the consecutive-miss deletion threshold E. In addition, the placement of the two horizontal boundary lines that define Regions A, B, and C is not specified, nor is the feature crop size beyond the mention of (120,120,3). Without these parameter values, the experimental results cannot be reproduced or compared with alternative settings.
[Section II-A] The fine-tuning details for the two SSD MobileNet head detectors are absent. The manuscript says the models are 'self-trained' and fine-tuned to distinguish heads from chairs, trolleys, and bags, but it does not report the size or composition of the training dataset, the annotation procedure, the number of training iterations, or any detection-level accuracy on a held-out test set. The examples in Figure 1 are illustrative but not a quantitative evaluation. This omission weakens the generalization claim for 'any complex environment.'
[Table III] The state-of-the-art comparison in Table III is not reliable. The baseline 'Chinthaka et al. [9]' shares two co-authors with the present paper (B. K. K. Ng and C. Yuen), and the accuracy/FPS values for the baselines are cited from prior publications with, presumably, different hardware, camera setups, and evaluation protocols. The claimed 2% improvement over the best baseline is therefore not an independent, matched comparison. The table should state the source and conditions of each baseline result and discuss whether the comparisons are apples-to-apples.

minor comments (4)

[Section II, first paragraph] There is a typo: 'we introduce the the proposed object detection model' should read 'we introduce the proposed object detection model.'
[Section II-D] Equation (2) defines accuracy as (Total observations - Error) / Total observations, but the term 'Error' is never defined precisely. Clarify whether an error is a wrong in/out count, a missed count, or any discrepancy per event.
[Section III, Table I] The FPS results in Table I are averages without standard deviations or the number of runs; adding this information would make the real-time claim more credible.
[Figure 3] The text references a real-time deployment result via link [12], but the figure caption does not explain the on-screen overlay ('# of Ins') or the ID labels A and C. A brief explanation in the caption would improve readability.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the accuracy claims are empirical measurements, and the only self-citation ([9]) is a non-load-bearing baseline.

full rationale

The paper is an empirical system paper, not a derivation, so most circularity patterns do not apply. The counting rule in Section II-D (entry = A to C, exit = C to A) is a deliberate operational definition of the target event, while accuracy in Eq. 2 is measured against external ground-truth observations, so the reported result is not its input by construction. Algorithm 1 contains thresholds T, D, and E, but no headline accuracy is claimed to be derived from fitting those thresholds to the test set, so there is no fitted-input-called-prediction step. The abstract's 97% conflicts with Table II's 92%, and the live test statistics are not reported; these are reproducibility and verification weaknesses, not circularity. The only self-citation is [9] (Chinthaka et al.), co-authored by B.K.K. Ng and C. Yuen, used as a SOTA baseline in Table III; the current accuracy is directly measured, and other baselines [7] and [8] are external, so the self-citation is not load-bearing and does not reduce the central claim to a citation chain. No specific equation-to-equation reduction to the paper's own inputs can be exhibited.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entity; the main burden falls on unstated tracker thresholds, the assumed monotonic crossing behavior, and undocumented training and test data. No code or dataset is provided.

free parameters (5)

Feature-distance threshold T
Algorithm 1 uses T to decide whether a new detection matches a tracked object; the paper never gives its value or how it was chosen.
Spatial-distance threshold D
Algorithm 1 rejects matches farther than D; D is not specified in the paper.
Miss-count threshold E
An object is removed after E consecutive missed detections (Algorithm 1, lines 15-18); E is not specified.
Region boundary line positions
The two horizontal lines dividing the field of view into Regions A, B, and C are chosen by the experimenters; their placement is not described.
Feature input crop size = 120x120x3
The input to MobileNetV2 is downsampled to (120, 120, 3) to reduce computation; no accuracy trade-off analysis is given.

assumptions (5)

domain assumption SSD MobileNet can detect heads from overhead cameras with sufficient accuracy after fine-tuning
Section II-A selects SSD MobileNet for its speed and adaptability but provides no detection accuracy evaluation on a held-out set.
domain assumption MobileNetV2 embeddings pretrained on ImageNet are discriminative for matching the same overhead head across frames
Section II-B relies on these embeddings for tracking without any evaluation of same-person versus different-person embedding distances.
domain assumption IR mode implies R=G=B for every pixel and is a reliable day/night switch
Section II-A uses Eq. 1 to decide which detection model to run; this pixel equality is a property of the specific camera and may not hold generally.
ad hoc to paper Every counted person crossing the field of view moves monotonically from A to C or C to A and remains tracked throughout
Section II-D does not define how partial or oscillating crossings are counted, so the head-count result depends on this unstated behavioral assumption.
domain assumption The live test ground truth is correct and the error term in Eq. 2 is meaningful
The accuracy formula uses an undefined 'Error' term and no ground-truth collection protocol is described.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Real-Time AI-Driven People Tracking and Counting Using Overhead Cameras." pith.science (2026). https://pith.science/paper/PVNOOKA6

@misc{pith2026241110072,
  author       = {Pith},
  title        = {Pith review of: Real-Time AI-Driven People Tracking and Counting Using Overhead Cameras},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PVNOOKA6}},
  note         = {Machine review of arXiv:2411.10072}
}

read the original abstract

Accurate people counting in smart buildings and intelligent transportation systems is crucial for energy management, safety protocols, and resource allocation. This is especially critical during emergencies, where precise occupant counts are vital for safe evacuation. Existing methods struggle with large crowds, often losing accuracy with even a few additional people. To address this limitation, this study proposes a novel approach combining a new object tracking algorithm, a novel counting algorithm, and a fine-tuned object detection model. This method achieves 97% accuracy in real-time people counting with a frame rate of 20-27 FPS on a low-power edge computer.

Figures

Figures reproduced from arXiv: 2411.10072 by the authors.

**Figure 1.** Before specifically training our model to identify potential disturbances, there were few mismatches, identifying distractions [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** This hand-crafted tracking approach offers a balance [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Real time deployment result can be found here [12]: (a) before entering the room. “# of Ins” is 9. (b) After two people, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 11 canonical work pages

[9]

Cluster pruning: An efficient filter pruning method for edge ai vision applica- tions,

C. Gamanayake, L. Jayasinghe, B. K. K. Ng, and C. Yuen, “Cluster pruning: An efficient filter pruning method for edge ai vision applica- tions,” IEEE Journal of Selected Topics in Signal Processing , vol. 14, no. 4, pp. 802–816, 2020

work page 2020
[1]

Security door system using human tracking method with laser range finders,

J. H. Lee, Y .-S. Kim, B. K. Kim, K. Ohba, H. Kawata, A. Ohya et al. , “Security door system using human tracking method with laser range finders,” in 2007 International Conference on Mechatronics and Automation. IEEE, 2007, pp. 2060–2065

work page 2007
[2]

Bi-directional passing people counting system based on ir-uwb radar sensors,

J. W. Choi, X. Quan, and S. H. Cho, “Bi-directional passing people counting system based on ir-uwb radar sensors,” IEEE Internet of Things Journal, vol. 5, no. 2, pp. 512–522, 2017

work page 2017
[3]

Door-monitor: Counting in-and-out visitors with cots wifi devices,

Y . Yang, J. Cao, X. Liu, and X. Liu, “Door-monitor: Counting in-and-out visitors with cots wifi devices,” IEEE Internet of Things Journal , vol. 7, no. 3, pp. 1704–1717, 2019

work page 2019
[4]

A people-counting and speed- estimation system using wi-fi signals,

L. Tian, L. Chen, Z. Xu, and Z. Chen, “A people-counting and speed- estimation system using wi-fi signals,” Sensors, vol. 21, no. 10, p. 3472, 2021

work page 2021
[5]

Counting people by using a single camera without calibration,

J. Cao, L. Sun, M. G. Odoom, F. Luan, and X. Song, “Counting people by using a single camera without calibration,” in 2016 Chinese control and decision conference (CCDC) . IEEE, 2016, pp. 2048–2051

work page 2016
[6]

Real-time top-view people counting based on a kinect and nvidia jetson tk1 integrated platform,

G. Li, P. Ren, X. Lyu, and H. Zhang, “Real-time top-view people counting based on a kinect and nvidia jetson tk1 integrated platform,” in 2016 IEEE 16th international conference on data mining workshops (ICDMW). IEEE, 2016, pp. 468–473

work page 2016
[7]

Benchmark data and method for real-time people counting in cluttered scenes using depth sensors,

S. Sun, N. Akhtar, H. Song, C. Zhang, J. Li, and A. Mian, “Benchmark data and method for real-time people counting in cluttered scenes using depth sensors,” IEEE Transactions on Intelligent Transportation Systems, vol. 20, no. 10, pp. 3599–3612, 2019

work page 2019

Show all 12 references

[8]

Passenger flow estimation based on convolutional neural network in public transportation system,

G. Liu, Z. Yin, Y . Jia, and Y . Xie, “Passenger flow estimation based on convolutional neural network in public transportation system,” Knowledge-Based Systems , vol. 123, pp. 102–115, 2017

2017
[10]

Ssd: Single shot multibox detector,

W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg, “Ssd: Single shot multibox detector,” in Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14 . Springer, 2016, pp. 21–37

2016
[11]

Mobilenets: Efficient convo- lutional neural networks for mobile vision applications,

A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient convo- lutional neural networks for mobile vision applications,” arXiv preprint arXiv:1704.04861, 2017

2017 arXiv
[12]

Real time testing video available online,

“Real time testing video available online,” https://bit.ly/4aqeWib, 2023

2023

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [9]

Cluster pruning: An efficient filter pruning method for edge ai vision applica- tions,

C. Gamanayake, L. Jayasinghe, B. K. K. Ng, and C. Yuen, “Cluster pruning: An efficient filter pruning method for edge ai vision applica- tions,” IEEE Journal of Selected Topics in Signal Processing , vol. 14, no. 4, pp. 802–816, 2020

work page 2020

[2] [1]

Security door system using human tracking method with laser range finders,

J. H. Lee, Y .-S. Kim, B. K. Kim, K. Ohba, H. Kawata, A. Ohya et al. , “Security door system using human tracking method with laser range finders,” in 2007 International Conference on Mechatronics and Automation. IEEE, 2007, pp. 2060–2065

work page 2007

[3] [2]

Bi-directional passing people counting system based on ir-uwb radar sensors,

J. W. Choi, X. Quan, and S. H. Cho, “Bi-directional passing people counting system based on ir-uwb radar sensors,” IEEE Internet of Things Journal, vol. 5, no. 2, pp. 512–522, 2017

work page 2017

[4] [3]

Door-monitor: Counting in-and-out visitors with cots wifi devices,

Y . Yang, J. Cao, X. Liu, and X. Liu, “Door-monitor: Counting in-and-out visitors with cots wifi devices,” IEEE Internet of Things Journal , vol. 7, no. 3, pp. 1704–1717, 2019

work page 2019

[5] [4]

A people-counting and speed- estimation system using wi-fi signals,

L. Tian, L. Chen, Z. Xu, and Z. Chen, “A people-counting and speed- estimation system using wi-fi signals,” Sensors, vol. 21, no. 10, p. 3472, 2021

work page 2021

[6] [5]

Counting people by using a single camera without calibration,

J. Cao, L. Sun, M. G. Odoom, F. Luan, and X. Song, “Counting people by using a single camera without calibration,” in 2016 Chinese control and decision conference (CCDC) . IEEE, 2016, pp. 2048–2051

work page 2016

[7] [6]

Real-time top-view people counting based on a kinect and nvidia jetson tk1 integrated platform,

G. Li, P. Ren, X. Lyu, and H. Zhang, “Real-time top-view people counting based on a kinect and nvidia jetson tk1 integrated platform,” in 2016 IEEE 16th international conference on data mining workshops (ICDMW). IEEE, 2016, pp. 468–473

work page 2016

[8] [7]

Benchmark data and method for real-time people counting in cluttered scenes using depth sensors,

S. Sun, N. Akhtar, H. Song, C. Zhang, J. Li, and A. Mian, “Benchmark data and method for real-time people counting in cluttered scenes using depth sensors,” IEEE Transactions on Intelligent Transportation Systems, vol. 20, no. 10, pp. 3599–3612, 2019

work page 2019