Monitoring of people entering and exiting private areas using Computer Vision

P Nagabhushan; Vinay Kumar V

REVIEW 5 major objections 6 minor 13 references

Monitoring of people entering and exiting private areas using Computer Vision

T0 review · 5 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper establishes that two outward-facing cameras can detect who enters and exits a camera-forbidden private area, introducing the EnEx2 benchmark with entry-exit accuracy above 96 percent.

desk verdict A useful but thin two-camera dataset with an overclaimed advantage: the paper does not actually show that the second camera helps. read the letter →

arxiv 1908.00716 v2 pith:7ANLRVGW submitted 2019-08-02 cs.CV

Vinay Kumar V , P Nagabhushan This is my paper · ORCID

classification cs.CV

keywords entry-exitsurveillanceprivateareastwo-cameradataseteventdetectionpeopletrackingpersonre-identificationEnEx2camera-forbidden

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

The paper is trying to establish that security cameras can monitor people entering and exiting private areas such as toilets and changing rooms without recording what happens inside, by placing two cameras outside and tracking each person until they cross the doorway. To make this measurable, it introduces EnEx2, a two-camera dataset of laboratory door crossings, with auto-generated tracks provided as ground truth. It then proposes a spatial transition rule in which an event is classified as Entry, Exit, or Just-appeared depending on whether the first and last bounding boxes of a person's track lie inside a manually marked entrance rectangle. On EnEx2 the rule reaches 96.2 percent Entry, 98.1 percent Exit, and 98.7 percent Just-appeared accuracy, which the paper takes as evidence that a second camera helps entry-exit surveillance.

What carries the argument

The central object is the entrance rectangle $E$, a manually marked region in the camera view that stands for the doorway of the private area, together with the subset test $B \subset E$ applied to a track's first and last bounding boxes. This test converts tracking into event classification. It is supported by a pipeline of background subtraction, HOG-based people detection, and Kalman-filter tracking that generates the tracks, and by the two-camera EnEx2 setup whose spatial coordinates are calibrated so that the opposite view can recover individuals occluded at the doorway.

What would settle it

Take a held-out set of EnEx2 clips where people overlap at the doorway, have human annotators label each appearance as Entry, Exit, or Just-appeared, and compare their labels with those produced by the bounding-box-subset rule; systematic disagreement on those clips would show that the reported accuracy depends on the auto-generated tracks rather than on the rule itself.

Watch

Extended reading notes

Core claim

The central claim is that entry-exit surveillance can be reduced to a geometric test on track endpoints. For each track, the initial bounding box state is classified as EA (entered the scene from outside) or EX (exited from the private area) by checking whether the box is a subset of the entrance rectangle E; the final state is classified as EN (entered the private area) or EA (exited the scene to outside) in the same way. The ordered pair of states then determines the event: (EA, EA) is Just appeared, (EA, EN) is Entry, (EX, EA) is Exit, and (EX, EN) is Re-entry, also treated as Just appeared. On the new EnEx2 dataset, with two oppositely placed cameras, this rule yields 96.2 percent Entry, 98.1 percent Exit, and 98.7 percent Just-appeared accuracy, above the single-camera EnEx baseline, which the paper attributes to the second view resolving doorway occlusions.

Load-bearing premise

The load-bearing premise is that whenever someone crosses the doorway, the people detector and tracker produce a complete, correctly placed bounding box in exactly the frame of the crossing, because the event label is simply whether that box falls inside the manually marked entrance rectangle.

Editorial extensions

If this is right

If the two-camera advantage holds, future entry-exit surveillance systems can rely on exterior cameras only, preserving privacy inside the private area while still logging who enters and exits.
EnEx2 provides a benchmark that researchers can use to compare entry-exit event detection and person re-identification after a privacy interval.
The spatial-transition rule is dataset-agnostic: the paper reports that the same rule also works on CAVIAR corridor sequences and PAMELA-UANDES metro boarding sequences by treating shop doors and train doors as entrance rectangles.
Entry-exit event detection is an extension of tracking, so improvements in people detection and tracking at doorways should directly improve event accuracy.

Reading between the lines

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

The event rule reads only the first and last frames of a track, so a natural extension is to classify events from the whole trajectory relative to the entrance rectangle, which could survive momentary detector failures at the doorway.
Because the two camera views are spatially calibrated, EnEx2 could also serve as a testbed for multi-view person re-identification and track association, not just event detection.
For real deployments the system stores no imagery from inside the private area, so the remaining privacy exposure is the exterior cameras themselves; a useful extension would add a privacy metric for bystanders caught by the outside view.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

First, the paper gives you something real: EnEx2, a two-camera entry-exit surveillance dataset with track-level ground truth, plus baseline numbers for a straightforward containment-based event detector. That is a useful step for a small subfield. The authors also run their method on EnEx, CAVIAR, and PAMELA-UANDES, which gives some comparability. The idea itself is not deep, but the dataset is the contribution.

The soft spots are significant. The ground truth for EnEx2 is pseudo-annotated using HOG detections plus background subtraction and Kalman filtering (Section 2.1), and evaluation uses the same HOG-based detector (Section 4). So the people detection and event detection scores partly measure the system against its own output. That is a circularity problem the paper never addresses. On top of that, Tables 1 and 2 report single numbers with no error bars, no track counts, no confusion matrices, so we have no sense of variance or where failures happen.

The bigger issue is the conclusion: 'High Entry-Exit event detection accuracy on EnEx2 dataset demonstrates the importance of a second camera.' That is not supported. EnEx2 is a different scene, different camera setup, different annotation procedure, and it is compared against EnEx which is single-camera. There is no matched experiment with the same scene and subjects in single- versus two-camera mode. The accuracy gaps in Table 2 are a couple of points, well within noise given the lack of error bars. The third column 'Just appeared' is basically the same. So the title claim overreaches.

A smaller point: EnEx2 is not released, so the dataset's utility is currently hypothetical. We cannot verify the pseudo-annotations or rerun the numbers.

All that said, the paper is not incoherent. The spatial transition rule is simple and the four-state formulation is clear. The authors are honest that it is pseudo-annotated. It deserves a serious referee, but only with major revision: release the data or at least independent annotations, add error analysis, and either run a matched single-vs-two-camera ablation or remove the causal claim about the second camera.

Who is this for? Researchers working on entry-exit surveillance or privacy-respecting monitoring of doorways. For that niche, the dataset is potentially useful if made available. For general CV, not much. I'd send it to peer review with a request for major revision rather than desk-rejecting it flat.

Referee Report

5 major / 6 minor

Summary. The paper introduces EnEx2, a two-camera video dataset for Entry-Exit Surveillance, and a simple rule-based event detection method that labels tracks as Entry, Exit, or Just appeared based on whether the first and last bounding boxes lie within a manually marked entrance rectangle. Baseline people-detection and event-detection results are reported for EnEx2 and three existing datasets (EnEx, CAVIAR, PAMELA-UANDES). The paper concludes that the high accuracy on EnEx2 demonstrates the importance of using a second camera.

Significance. If the dataset were released and independently validated, EnEx2 could serve as a useful benchmark for privacy-respecting monitoring of private-area doorways, and the paper addresses a real application gap. The paper is transparent about EnEx2 being pseudo-annotated and makes a clear statement of the problem and the proposed event taxonomy. However, the reported results are not sufficient to establish the central claim about camera count, and the lack of dataset release and independent annotation severely limits the contribution's immediate utility.

major comments (5)

[Section 5, Table 2] The conclusion that high accuracy on EnEx2 demonstrates the importance of a second camera is unsupported. Table 2 compares EnEx (single camera, different location, different annotation procedure) with EnEx2 (two cameras, different scene, pseudo-annotations); there is no matched experiment that runs the same scene, subjects, detector, tracker, and annotation protocol in single-camera versus two-camera modes. The reported differences (Entry 96.2 vs 94.4, Exit 98.1 vs 96.4, Just appeared 98.7 vs 98.3) are small, and with no error bars or significance tests they cannot be attributed to camera count.
[Section 2.1 and Section 4] The EnEx2 ground-truth tracks are generated using the same HOG-based people detector [6] (with background subtraction [5] and Kalman filters [8]) that is later used in the evaluation pipeline. This creates a circularity: the reported EnEx2 detection and event-detection scores in Tables 1 and 2 partly measure how well the detector agrees with its own pseudo-labels rather than how well it matches independent ground truth. The paper labels the dataset 'pseudo-annotated' but provides no manual validation or independent test set to break the circularity.
[Section 3.2, Table 2] The event detection method relies on the assumption that the first and last bounding boxes of each track are correctly detected and that the entrance rectangle E is accurately marked. No analysis is provided for cases of occlusion, track fragmentation, or bounding-box drift near the entrance, which can change the B ⊂ E decision and hence the event label. The two free parameters (E coordinates and the unassigned-track frame threshold) are not specified in the paper, so the reported numbers cannot be reproduced.
[Section 4] The evaluation methodology is under-specified. Tables 1 and 2 give single performance numbers without error bars, track counts, confusion matrices, or a definition of the reported accuracy (e.g., per-track or per-frame). It is unclear how many video sequences from each dataset were used, how events were defined, and whether the results are averaged over runs. Without this information, the results cannot be assessed statistically or reproduced.
[Section 1.2 and Section 5] The paper's main deliverable—the EnEx2 dataset—is not released; no URL, download link, or access instructions appear. As a benchmark contribution, the dataset must be available for independent verification and comparison. Additionally, the claim that two cameras improve re-identification accuracy is not evaluated at all; the paper only reports detection and event-detection results, not re-identification results, despite listing re-identification as a motivation.

minor comments (6)

[Section 2.1] The paper contains several incomplete references to figures: 'as shown in Figure .' and 'the view in the figure 3b' do not correspond to any labeled figure in the text.
[Section 3.2] The notation for bounding box inclusion is inconsistent: sometimes it appears as 'B⊂ E' and sometimes as 'B⊂E'; please standardize the math notation.
[Section 4] The term 'deeply learned people detection method' for HOG is inaccurate; HOG features are not deep learning.
[Table 1 and Section 4] The dataset name is written inconsistently: 'CA VIAR' in Table 1 vs 'CAVIAR' in Section 4. Please unify the spelling.
[References] The reference list contains errors: reference [2] lists the author as 'V , Vinay Kumar' instead of 'Vinay Kumar V', and reference [12] has a missing space in 'PAMELA- UANDES'.
[Abstract and Section 5] The term 'pseudo-annotated' is used without explaining what pseudo-annotation entails; given the circularity concern, this should be clarified in the methodology.

Circularity Check

1 steps flagged · score 6.0 of 10

EnEx2's ground-truth tracks are generated by the same HOG/Kalman pipeline used for evaluation, so the dataset's headline accuracies are partly self-agreement; the second-camera conclusion is confounded by a cross-dataset comparison.

fitted input called prediction [Section 2.1 (EnEx2 dataset construction) and Section 4 (Evaluation; Tables 1-2)]
"Tracks of every individual in the video sequences with their positions in each frame are provided as ground truth. ... Tracks are built using intelligent people detectors [6] after subtracting frames from the background [5]. ... Given the video stream as input to the system, individuals are detected in every frame using Histogram of Oriented Gradient based deeply learned people detection method [6]."

The EnEx2 ground truth is pseudo-annotated by the paper's own pipeline: the same HOG detector [6], background subtraction [5], and Kalman filtering [8] that create the dataset tracks are reused in Section 4 to produce the detections and tracks fed into the event detector. People-detection accuracy on EnEx2 thus compares the detector with boxes generated by that detector, so it is partly a self-consistency score rather than an independent prediction. Additionally, Section 3.2 defines entry/exit events by checking whether a track's first/last bounding box lies inside the entrance rectangle E; because the EnEx2 event labels are derived from the same auto-generated tracks, the event-detection accuracy on EnEx2 substantially reapplies the labeling rule to its own annotations.

full rationale

The paper is mostly a dataset-and-baseline contribution, and there is no deep derivation chain to audit. The main circularity is localized to EnEx2: its ground-truth tracks are pseudo-annotations produced by the same HOG detector, background-subtraction, and Kalman-tracker pipeline that Section 4 uses at evaluation time. Consequently, the EnEx2 people-detection and event-detection numbers in Tables 1 and 2 partly measure the system's agreement with labels its own pipeline generated, and the Section 3.2 event rule (first/last bounding box inside the marked entrance rectangle) is also the natural rule used to create pseudo-event-labels from those tracks. The high EnEx2 event accuracy is therefore substantially forced by construction, and Section 5 directly uses that accuracy to conclude that a second camera is important. The evaluations on EnEx, CAVIAR, and PAMELA-UANDES provide some independent content, but Table 2 is not a matched single-camera versus two-camera ablation; it compares different scenes, different annotation processes, and different camera geometries, and the reported gaps are small with no error bars or significance tests. The self-citations to the authors' prior EnEx dataset are present but are not themselves the load-bearing circular step; the circularity is the shared pseudo-annotation and evaluation pipeline.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central pipeline assumes that bounding-box containment in a fixed entrance rectangle identifies entry or exit, that detector and tracker outputs are reliable, and that pseudo-annotated tracks can serve as ground truth. Two hand-chosen parameters, the entrance region and the missing-track threshold, are not specified. No new physical or conceptual entities are introduced.

free parameters (2)

Entrance rectangle E coordinates = not specified
The event classifier decides Entry or Exit solely by containment of the track's first or last bounding box in this manually marked rectangle (Section 3.2). No coordinates, size, or marking protocol is reported.
Unassigned-track frame threshold = not specified
Track endpoints are determined by 'unassigned after a thresholded number of frames' (Section 4). The threshold is never quantified and directly controls whether a person is labeled as entering the private area or as a missing track.

assumptions (3)

domain assumption Every person entering or exiting the private area is detected and tracked as one continuous bounding-box track, with track endpoints exactly at the moments of crossing the entrance.
Event labels in Section 3.2 are defined entirely from the first and last frames of each track. If detection or tracking fails, the endpoint containment test will mislabel events.
domain assumption The entrance of the private area can be represented as a fixed 2D rectangle in the camera view.
Section 3.2 defines E as a rectangle and uses the subset test B subset E. An oblique or perspective view of a doorway cannot always be represented accurately by a single axis-aligned rectangle.
domain assumption Auto-generated tracks from the people detector [6], background subtraction [5], and Kalman filters [8] are valid ground truth.
Section 2.1 says tracks are built using these components and provided as ground truth, with no manual verification or inter-annotator agreement described.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Monitoring of people entering and exiting private areas using Computer Vision." pith.science (2026). https://pith.science/paper/7ANLRVGW

@misc{pith2026190800716,
  author       = {Pith},
  title        = {Pith review of: Monitoring of people entering and exiting private areas using Computer Vision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7ANLRVGW}},
  note         = {Machine review of arXiv:1908.00716}
}

read the original abstract

Entry-Exit surveillance is a novel research problem that addresses security concerns when people attain absolute privacy in camera forbidden areas such as toilets and changing rooms that are basic amenities to the humans in public places such as Shopping malls, Airports, Bus and Rail stations. The objective is, if not inside these camera forbidden areas, from outside, the individuals are to be monitored to analyze the time spent by them inside and also the suspecting transformations in their appearances if any. In this paper, firstly, a pseudo-annotated dataset of a laboratory observation of people entering and exiting the camera forbidden area captured using two cameras in contrast to the state-of-the-art single-camera based EnEx dataset is presented. Conventionally the proposed dataset is named \textbf{\textit{EnEx2}}. Next, a spatial transition based event detection to determine the entry or exit of individuals is presented with standard results by evaluating the proposed model using the proposed dataset and the publicly available standard video surveillance datasets that are hypothesized to Entry-Exit surveillance scenarios. The proposed dataset is expected to enkindle active research in Entry-Exit Surveillance domain.

Figures

Figures reproduced from arXiv: 1908.00716 by the authors.

**Figure 2.** Illustration of transition of individuals from state EA to EN Let there be n individuals who appears in the camera view scene in the interval [tStart, tEnd] be represented by S = {S1, S2, ..., Sn}. For each individual, before creating a track, it is important to determine whether the individual entered the scene from outside or exited from private area. If the individual detected as a new track is bounded in a bound… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 11 canonical work pages

[6]

Fast Feature Pyramids for Object Detection

Dollar, P., R. Appel, S. Belongie, and P. Perona. "Fast Feature Pyramids for Object Detection." IEEE Transactions on Pattern Analysis and Machine Intelligence. Vol. 36, Issue 8, 2014, pp. 15321545

work page 2014
[5]

Dalal and B

N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In CVPR, 2005

2005
[8]

Region-of-Interest Tracking Based on Keypoint Trajectories on a Group of Pictures,

V. Garcia, E. Debreuve and M. Barlaud, "Region-of-Interest Tracking Based on Keypoint Trajectories on a Group of Pictures," 2007 International Workshop on Content-Based Multimedia Indexing, Bordeaux, 2007, pp. 198-203

work page 2007
[1]

Dufour, Jean-Yves, Intelligent Video Surveillance Systems, John Wiley Publisher (2012)

work page 2012
[2]

In: Chaudhuri, B.B., Nakagawa, M., Khanna, P., Kumar, S

V, Vinay Kumar, Nagabhushan,P, (2019) Entry-Exit Video Surveillance: A benchmark dataset. In: Chaudhuri, B.B., Nakagawa, M., Khanna, P., Kumar, S. (Eds.) Proceedings of 3rd International Conference on Computer Vision & Image Processing. Advances in Intelligent Systems and Computing, vol 1022. Springer, Singapore

work page 2019
[3]

Entry-Exit event detection from video frames

Vinay Kumar V, P Nagabhushan, “Entry-Exit event detection from video frames”, International Journal of Computer Sciences and Engineering, Vol.6, Issue.2, pp.112-118, 2018

work page 2018
[4]

(2005) A Novel Adaptive Gaussian Mixture Model for Background Subtraction

Cheng J., Yang J., Zhou Y. (2005) A Novel Adaptive Gaussian Mixture Model for Background Subtraction. In: Marques J.S., Prez de la Blanca N., Pina P. (eds) Pattern Recognition and Image Analysis. IbPRIA 2005. Lecture Notes in Computer Science, vol 3522. Springer, Berlin, Heidelberg

work page 2005
[7]

University of North Carolina at Chapel Hill, Department of Computer Science

Welch, Greg, and Gary Bishop, An Introduction to the Kalman Filter, TR 95041. University of North Carolina at Chapel Hill, Department of Computer Science

work page

Show all 13 references

[9]

Adaptive background mixture models for real-time tracking,

C. Stauffer and W. E. L. Grimson, "Adaptive background mixture models for real-time tracking," Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149), Fort Collins, CO, USA, 1999, pp. 246-252 Vol. 2. doi: 10.1109/CVPR.19...

1999
[10]

R. B. Fisher, PETS04 Surveillance Ground Truth Data Set, Proc. Sixth IEEE Int. Work. on Performance Evaluation of Tracking and Surveillance (PETS04), pp 1-5, May 2004

2004
[11]

5 (6 pp.)-5 (6 pp.), DOI: 10.1049/cp.2018.1281

Belloc, M.; Velastin, S.A.; Fernandez, R.; Jara, M.: 'Detection of People Boarding/Alighting a Metropolitan Train using Computer Vision', IET Conference Proceedings, 2018, p. 5 (6 pp.)-5 (6 pp.), DOI: 10.1049/cp.2018.1281

2018
[12]

Video Object Segmentation and Tracking: A Survey

Rui Yao, Guosheng Lin, Shixiong Xia, Jiaqi Zhao, Yong Zhou.2019. Video Object Segmentation and Tracking: A Survey. arXiv preprint arXiv:1904.09172 (2019)

2019 arXiv
[13]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [6]

Fast Feature Pyramids for Object Detection

Dollar, P., R. Appel, S. Belongie, and P. Perona. "Fast Feature Pyramids for Object Detection." IEEE Transactions on Pattern Analysis and Machine Intelligence. Vol. 36, Issue 8, 2014, pp. 15321545

work page 2014

[2] [5]

Dalal and B

N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In CVPR, 2005

2005

[3] [8]

Region-of-Interest Tracking Based on Keypoint Trajectories on a Group of Pictures,

V. Garcia, E. Debreuve and M. Barlaud, "Region-of-Interest Tracking Based on Keypoint Trajectories on a Group of Pictures," 2007 International Workshop on Content-Based Multimedia Indexing, Bordeaux, 2007, pp. 198-203

work page 2007

[4] [1]

Dufour, Jean-Yves, Intelligent Video Surveillance Systems, John Wiley Publisher (2012)

work page 2012

[5] [2]

In: Chaudhuri, B.B., Nakagawa, M., Khanna, P., Kumar, S

V, Vinay Kumar, Nagabhushan,P, (2019) Entry-Exit Video Surveillance: A benchmark dataset. In: Chaudhuri, B.B., Nakagawa, M., Khanna, P., Kumar, S. (Eds.) Proceedings of 3rd International Conference on Computer Vision & Image Processing. Advances in Intelligent Systems and Computing, vol 1022. Springer, Singapore

work page 2019

[6] [3]

Entry-Exit event detection from video frames

Vinay Kumar V, P Nagabhushan, “Entry-Exit event detection from video frames”, International Journal of Computer Sciences and Engineering, Vol.6, Issue.2, pp.112-118, 2018

work page 2018

[7] [4]

(2005) A Novel Adaptive Gaussian Mixture Model for Background Subtraction

Cheng J., Yang J., Zhou Y. (2005) A Novel Adaptive Gaussian Mixture Model for Background Subtraction. In: Marques J.S., Prez de la Blanca N., Pina P. (eds) Pattern Recognition and Image Analysis. IbPRIA 2005. Lecture Notes in Computer Science, vol 3522. Springer, Berlin, Heidelberg

work page 2005

[8] [7]

University of North Carolina at Chapel Hill, Department of Computer Science

Welch, Greg, and Gary Bishop, An Introduction to the Kalman Filter, TR 95041. University of North Carolina at Chapel Hill, Department of Computer Science

work page