Pith. sign in

REVIEW 3 major objections 4 minor 56 references

Unsupervised Detection of Entry and Exit Regions from Vehicle Trajectories for Camera-Agnostic Turning Movement Counts

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that entry and exit regions for turning movement counts can be learned from raw vehicle trajectories alone, with no labels or camera calibration, achieving a median classification error of 3.4% across 25 uncalibrated surve

desk verdict A genuinely useful unsupervised TMC region-identification pipeline with a solid engineering evaluation, but the headline 3.4% is likely computed on the same trajectories used to estimate the regions, so the forward-classification evidence rests mainly on the 9-camera temporal-split experiments. read the letter →

arxiv 2607.10949 v2 pith:335D5VXB submitted 2026-07-12 cs.CV

Parikshit Singh Rathore , Vishwajeet Pattanaik , Punit Rathore This is my paper
classification cs.CV
keywords turningmovementcountsentry/exitregiondetectionunsupervisedlearningvehicletrajectoriestrajectoryclusteringcamera-agnosticsurveillancecamerastrafficflowanalysis
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 tries to show that the bottleneck blocking network-wide turning movement counts—manually drawing entry and exit zones on every camera—can be removed entirely. It proposes a six-step pipeline that clusters the first and last points of tracked vehicle trajectories into persistent spatial region polygons, then classifies future trajectories by point-in-polygon containment in linear time. On 25 Bengaluru traffic cameras, including 16 held-out feeds, the pipeline reaches a median classification error of 3.4% (median GEH 2.43), without any per-camera annotation, calibration, or prior knowledge of intersection geometry. The central insight is that trajectory endpoints, once cleaned and restricted to a peripheral buffer zone, carry enough boundary information to reconstruct entry and exit regions as reusable polygons.

What carries the argument

The load-bearing object is the exclusion zone: a concentric rectangle inset from the ROI by margins equal to one grid cell (ratio 1/n), which removes the central intermingling region where entry and exit flows overlap in the image plane and restricts clustering to a peripheral buffer. This zone is coupled to the ROI estimated in Step 2—errors in the bounding rectangle propagate directly into zone placement—and to the n×n grid used for polygon construction. The second key mechanism is point selection: taking only the first s and last s positions of each trajectory (s=7) suppresses interior path information that dilutes cluster boundaries. Together these convert trajectory endpoints into spati

What would settle it

Take a camera where vehicles enter or exit through the middle of the frame—e.g., a side-mounted camera that frames an intersection with a driveway in the center—run the pipeline with the recommended configuration, and compare the recovered region polygons to manual annotations. If the peripheral-ring assumption holds, the polygons track the true entry and exit zones; if entry/exit flow lies near the scene interior, the exclusion zone carves it away and classification error rises far above the 3.4% median.

Watch

Extended reading notes

Core claim

The central claim is that entry and exit regions at an intersection can be recovered from the raw endpoints of detected-and-tracked vehicle trajectories alone, and that the recovered regions are persistent enough to classify future trajectories without re-clustering. The pipeline filters tracking artifacts, computes a region of interest as the bounding rectangle of trajectory coordinates, selects a small number of initial and terminal points per trajectory (s=7 recommended), enforces a peripheral buffer via a concentric exclusion zone, clusters the surviving points with K-Means, and encloses each cluster in grid-based polygons. The statistical analysis of 17,100 pipeline executions identifie

Load-bearing premise

The load-bearing premise is that genuine entry and exit points always occupy a peripheral ring around the bounding rectangle of observed vehicle motion, so that a concentric rectangular exclusion zone can separate them from the intersection interior; if vehicles enter or exit near the scene interior, or the ROI misplaces the intersection center, the whole separation collapses.

Editorial extensions

If this is right

  • Traffic agencies could generate turning movement counts from existing city surveillance networks without per-camera manual polygon specification, removing the main bottleneck to network-wide deployment.
  • Because classification is point-in-polygon containment rather than pairwise trajectory similarity, the approach scales to continuous monitoring of large camera fleets at O(N) per batch, where trajectory clustering baselines cost O(N^2).
  • The recommended configuration—s=7, K-Means, exclusion ratio 1/6, ROI on, grid polygons—provides a concrete default that transfers across heterogeneous traffic conditions and camera geometries.
  • Calibration clips of at least 60 minutes, ideally from peak-traffic periods, further improve region estimation, giving operational guidance for when to run the one-time setup.

Reading between the lines

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

  • If the peripheral-ring assumption generalizes, the same pipeline could be applied to other boundary-detection problems where sources and sinks are defined by trajectory endpoints—e.g., pedestrian flow in public spaces or entry/exit detection in warehouses—by treating the observed-motion bounding rectangle as the relevant frame.
  • The claim that K-Means outperforms GMM and DBSCAN because of implicit regularization is a hypothesis the paper states but does not isolate; it could be tested by adding controlled synthetic clusters with known noise levels.
  • The macro-averaged error metric masks difficulty on rare movements like U-turns; a volume-weighted or per-movement worst-case evaluation would sharpen the operational picture, and the authors acknowledge this in the limitations.
  • The paper's fixed preprocessing thresholds (e.g., δ_min=72 pixels) are tuned for 1920×1080; a resolution-normalized version would let the pipeline be dropped onto lower-resolution cameras without re-tuning—an extension the paper leaves to future work.
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.

Referee Report

3 major / 4 minor

Summary. The paper proposes an unsupervised, camera-agnostic pipeline for identifying entry/exit regions from raw vehicle trajectories and using those regions to classify turning movements. The pipeline consists of trajectory preprocessing, ROI estimation, endpoint selection (with optional density equalization), an exclusion zone, clustering, and polygon construction, followed by point-in-polygon classification. The authors evaluate 900 configurations over 9 Bengaluru development cameras and 10 UA-DETRAC sequences, use parametric and nonparametric tests to derive a recommended configuration, and report a median classification error of 3.42% and median per-movement GEH of 2.43 across 25 Bengaluru cameras, including 16 held-out locations. They also compare against two trajectory-clustering baselines and study the effect of calibration duration and traffic density.

Significance. If the headline results hold, this is a practically valuable contribution: it removes the per-camera manual annotation bottleneck for turning movement counts, operates without calibration, and produces persistent regions that classify new trajectories in O(N) rather than O(N^2). The evaluation is unusually broad for this problem: 17,100 pipeline executions, 25 real surveillance cameras, held-out locations, temporally independent video for parameter selection, GEH reporting, and baseline comparisons are all strengths. The systematic parameter analysis is also a useful methodological contribution. However, the evaluation protocol must demonstrate that the headline error measures forward classification of future trajectories, not in-sample cluster purity, and the statistical analysis needs to account for within-camera correlation. These issues are fixable and do not undermine the core idea, but they affect the strength of the central claim.

major comments (3)
  1. [Section IV-C and Section V-D, Fig. 3 (Eqs. 8–9)] The headline 3.42% median error is not established as a forward-classification result. Eq. (8) assigns each predicted polygon the ground-truth label by maximizing overlap with P_boundary, and Eq. (9) computes the error on T_filtered. Section V-D says the recommended configuration was applied to all 25 cameras using the temporally independent video, but does not state that the trajectories used to estimate and label regions are disjoint from those used to score error. If the same clip is used for both, the result measures in-sample cluster purity under an oracle label mapping, not the ability of persistent regions to classify future trajectories. Section VI supplies a temporal split for the 9 development cameras only; the 16 held-out locations are not temporally split. Please clarify the split, and if regions were estimated and evaluated on the same trajectories, re-evaluate on temporally
  2. [Section IV-D, Tables II–III] The one-way ANOVA analyses treat each of the 900 configurations per feed as an independent observation, pooling 17,100 executions across 19 feeds without modeling camera as a random effect or within-camera correlation. The p-values in Tables II and III are therefore anticonservative. The nonparametric tests partially address this, but they are reported selectively (omnibus tests and a limited set of post-hoc comparisons). The claim that the same three parameters are 'consistently significant' should be based on tests that account for feed-level correlation or explicitly justify why the within-feed correlations do not affect the conclusions.
  3. [Section V-C (synthesis) vs. Table IV] The second finding states that 'K-Means consistently outperforms GMM and DBSCAN, a result that is significant in every analysis conducted.' This is contradicted by Table IV: on Bengaluru, the Nemenyi comparison between K-Means and GMM does not reach significance (only K-Means and GMM are each significant versus DBSCAN). The recommended configuration may still be defensible, but the wording should be qualified to avoid overstating the consistency of the significance result.
minor comments (4)
  1. [Fig. 2 vs. Fig. 3] The labels 'train set' for Fig. 2 and 'test set' for Fig. 3 are confusing because the method is unsupervised and there is no conventional supervised training. 'Development cameras' and 'temporally independent evaluation' would be clearer.
  2. [Section IV-A] Please report the number of trajectories per camera and per clip. This is needed to interpret macro-averaged errors and the GEH percentages, particularly because the manuscript itself notes in Section VII that low-volume movements dominate the TN term and can inflate apparent accuracy.
  3. [Section VI-A, Fig. 6] The 180-minute calibration condition is plotted with a median of 1.37 versus 1.36 for 60 minutes; the text says improvement 'plateaus,' but it would be useful to also report a statistical comparison of these two conditions rather than only visual inspection.
  4. [Section IV-A] The statement that preprocessing was not applied to UA-DETRAC means the cross-dataset comparison is not apples-to-apples. This is acknowledged in the text, but the impact of this asymmetry on the parameter-sensitivity conclusions could be discussed more explicitly.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline 3.42% error is computed on the same trajectories used to fit the regions and to assign ground-truth labels, so the central 'prediction' partially reduces to an in-sample fit.

  1. fitted input called prediction [Section III (Turning Movement Classification); Section IV-C Eq. (8); Section V-D]
    "Once entry and exit regions have been delineated, the preprocessed trajectories from T_filtered are classified by assigning each trajectory's initial point to the entry region in which it falls and its terminal point to the corresponding exit region. ... For each predicted polygon P_i, the assigned label is that of the ground-truth polygon sharing the greatest number of initial or terminal trajectory points ... P_boundary is the set of boundary points defined in Step 4. ... the pipeline achieves a median classification error of 3.42% across all 25 camera locations."

    The regions are estimated from T_filtered, and the same T_filtered is then classified to compute the error. Equation (8) chooses each predicted polygon's ground-truth label by maximizing shared points in P_boundary — the same points used in Equation (9) to compute the trajectory assignment error. Section V-D states only that the video is temporally independent from the parameter-selection data; it does not state that the trajectories used to estimate/label regions are disjoint from the trajectories used to evaluate the error. Thus the 3.42% median error measures in-sample cluster purity under an optimal test-time label mapping, not forward classification of unseen trajectories. The '16 held-out locations' are new cameras, not held-out trajectories.

full rationale

The pipeline's region-estimation chain — preprocessing, ROI bounding, endpoint selection, exclusion zone, clustering, polygon construction — is self-contained and does not derive its output from the ground-truth labels. There are no load-bearing self-citations or imported uniqueness theorems. The only significant circularity is in the headline evaluation: the classification step runs on the same T_filtered used to estimate the regions, and the label association in Eq. (8) uses the same P_boundary points that are then scored in Eq. (9). Without an explicit trajectory-level calibration/test split for the 25-camera evaluation, the reported median error is an in-sample fit. Section VI does perform a proper temporal split for the nine development cameras, providing partial independent support, but the central 25-camera claim remains partly circular.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The pipeline's central claim rests on several domain assumptions rather than new math: the peripheral-buffer geometry of entry/exit regions, the sufficiency of a rectangular ROI/exclusion zone, automatic K selection recovering true region counts, and expert-annotated ground truth as an unbiased reference. No new entities are postulated.

free parameters (8)
  • δ_min (displacement threshold) = 72 px
    Step 1; set empirically for 1920×1080 and held fixed across all runs; not validated at other resolutions (stated in §VII-B).
  • τ_min (temporal filter threshold) = 1.5 s
    Step 1; chosen to remove identity-switch fragments; fixed across deployments.
  • r / n_min (neighborhood filter) = r=50 px, n_min=3
    Step 1; fixed thresholds with empirical justification in S-I.
  • s (points per end) = 7
    Recommended value selected on development cameras via ANOVA/Friedman; s∈{1,3,5,7,9,All} tested.
  • Exclusion ratio 1/n = 1/6
    Recommended value selected on development data; also sets grid resolution.
  • Clustering algorithm = K-Means
    Selected by statistical significance over GMM and DBSCAN on dev data.
  • Polygon representation = Grid
    Directional evidence on UA-DETRAC; adopted as default with stated weak support.
  • ROI estimation / density equalization = ROI enabled, subsampling disabled
    Defaults adopted from directional evidence or lack of significance.
assumptions (5)
  • domain assumption Entry/exit regions are recoverable from initial/terminal point locations after preprocessing; true entry/exit points lie in the peripheral buffer of the trajectory-bounding ROI
    Steps 3-4; the entire pipeline is built on this geometric premise.
  • domain assumption A rectangular ROI and a concentric rectangular exclusion zone can separate all entry/exit flows in the image plane
    Steps 2 and 4 (Eqs. 3-5); fails for cameras with interior entry points or non-convex layouts; no such cases reported.
  • domain assumption Automatic cluster-count selection (knee-point for K-Means/DBSCAN, BIC for GMM) within K∈[2,9] recovers the true number of entry/exit regions
    Step 5; intersections with more than 9 entry+exit regions cannot be represented.
  • domain assumption Expert-annotated ground-truth polygons and the many-to-one label association of Eq. 8 yield an unbiased evaluation of classification error
    Section IV-C; no inter-annotator agreement computed (acknowledged §VII-B); many-to-one mapping may hide over-segmentation.
  • domain assumption Upstream RT-DETR + BoT-SORT-ReID trajectories are accurate enough that region identification and classification are meaningful
    Section IV-A; the pipeline cannot compensate for severe occlusion; errors not disaggregated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unsupervised Detection of Entry and Exit Regions from Vehicle Trajectories for Camera-Agnostic Turning Movement Counts." pith.science (2026). https://pith.science/paper/335D5VXB

@misc{pith2026260710949,
  author       = {Pith},
  title        = {Pith review of: Unsupervised Detection of Entry and Exit Regions from Vehicle Trajectories for Camera-Agnostic Turning Movement Counts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/335D5VXB}},
  note         = {Machine review of arXiv:2607.10949}
}
read the original abstract

Turning movement counts are essential for intersection-level traffic management, yet their collection remains predominantly manual due to the cost of per-camera region annotation. This paper presents an unsupervised pipeline that identifies entry and exit regions directly from raw vehicle trajectories extracted via object detection and multi-object tracking, requiring no manual annotation, camera calibration, or prior knowledge of intersection geometry. Unlike trajectory clustering methods that classify individual trajectories using pairwise similarity and must be re-executed on every new batch, the proposed pipeline clusters initial and terminal point locations to produce persistent spatial region polygons that classify future trajectories by point-in-polygon containment at linear cost. The pipeline comprises six sequential steps, five of which introduce configurable parameters evaluated through a systematic statistical analysis spanning 17,100 pipeline executions across 9 surveillance cameras capturing dense heterogeneous traffic in Bengaluru, India, and 10 sequences from the UA-DETRAC benchmark dataset. Both parametric and nonparametric testing frameworks identify three consistently significant parameters and yield an empirically grounded recommended configuration. Under this configuration, the pipeline achieves a median classification error of 3.4% across all 25 Bengaluru cameras, including 16 held-out locations, with a median per-turning-movement GEH of 2.43. Compared with two trajectory clustering baselines, the proposed pipeline exhibits greater stability across camera views and lower computational cost, at the expense of higher median error. Extended evaluation demonstrates that calibration clips of at least 60 minutes and peak-traffic selection further improve region estimation quality.

Figures

Figures reproduced from arXiv: 2607.10949 by the authors.

Figure 1
Figure 1. Proposed pipeline for unsupervised entry and exit region identification. The six components progressively refine the raw [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 4
Figure 4. Distribution of GEH statistic values across all 25 [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 2
Figure 2. Classification error under the recommended configura [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Distribution of classification error across all 25 camera [PITH_FULL_IMAGE:figures/full_fig_p010_3.png]
Figure 5
Figure 5. Figure 5: Classification error distribution across all 25 Bengaluru [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 7
Figure 7. Figure 7: Distribution of classification error across the least busy, [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 6
Figure 6. Figure 6: Distribution of per-movement classification error across [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 1 linked inside Pith

  1. [1]

    Vehicle count- ing using computer vision: A survey,

    D. E. Valdivieso Tituana, S. G. Yoo, and R. O. Andrade, “Vehicle count- ing using computer vision: A survey,” in2022 IEEE 7th International conference for Convergence in Technology (I2CT), 2022, pp. 1–8

    2022

  2. [2]

    2012 national traffic signal report card: Technical report,

    National Transportation Operations Coalition (NTOC), “2012 national traffic signal report card: Technical report,” Institute of Transportation Engineers (ITE), Washington, D.C., Tech. Rep., 2012

    2012

  3. [3]

    Outcome-oriented performance measures for management of signalized arterial capacity,

    C. M. Day, J. R. Sturdevant, and D. M. Bullock, “Outcome-oriented performance measures for management of signalized arterial capacity,” Transportation Research Record: Journal of the Transportation Research Board, vol. 2192, no. 1, p. 24–36, Jan. 2010

    2010

  4. [4]

    Honeywell installs 7000 AI video cameras as a part of Bengaluru’s safe city project,

    Tech Desk, “Honeywell installs 7000 AI video cameras as a part of Bengaluru’s safe city project,” https://indianexpress.com/article/techno logy/techook/bengaluru-gets-over-7000-video-cameras-safe-city-proje cts-8524911/, 2023, accessed: 2026-04-19

    2023

  5. [5]

    New in the New Year — Safe City Project: 10,000 AI cameras with distress detection technology to watch over Delhi soon,

    A. Singh, “New in the New Year — Safe City Project: 10,000 AI cameras with distress detection technology to watch over Delhi soon,” https://indianexpress.com/article/cities/delhi/new-in-the-new-year-saf e-city-project-10000-ai-cameras-with-distress-detection-technology-t o-watch-over-delhi-soon-10437745/, 2025, accessed: 2026-04-19

    2025

  6. [6]

    Identifying vehicle turning movements at intersections from trajectory data,

    E. D. Saldivar-Carranza, H. Li, and D. M. Bullock, “Identifying vehicle turning movements at intersections from trajectory data,” in2021 IEEE International Intelligent Transportation Systems Conference (ITSC), 2021, pp. 4043–4050

    2021

  7. [7]

    Object detection in 20 years: A survey,

    Z. Zou, Z. Shi, Y . Guoet al., “Object detection in 20 years: A survey,” CoRR, vol. abs/1905.05055, 2019

    arXiv 1905

  8. [8]

    Object tracking: A survey,

    A. Yilmaz, O. Javed, and M. Shah, “Object tracking: A survey,”ACM Comput. Surv., vol. 38, no. 4, p. 13–es, Dec. 2006

    2006

Show all 56 references
  1. [9]

    Vision-based turning movement monitoring: Count, speed & waiting time estimation,

    M. Shokrolah Shirazi and B. T. Morris, “Vision-based turning movement monitoring: Count, speed & waiting time estimation,”IEEE Intelligent Transportation Systems Magazine, vol. 8, no. 1, pp. 23–34, 2016

    2016

  2. [10]

    Fast vehicle turning-movement count- ing using localization-based tracking,

    D. Gloudemans and D. B. Work, “Fast vehicle turning-movement count- ing using localization-based tracking,” in2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2021, pp. 4150–4159

    2021

  3. [11]

    Ground plane projection for improved traffic analytics at intersections,

    S. Pakdamansavoji, K. V . Jha, B. Abdulhaiet al., “Ground plane projection for improved traffic analytics at intersections,” in2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), 2024, pp. 741–748

    2024

  4. [12]

    Characterization of image distortions in multi- camera systems,

    M. Solh and G. AlRegib, “Characterization of image distortions in multi- camera systems,” inProceedings of the 2nd International Conference on Immersive Telecommunications, ser. IMMERSCOM ’09. Brussels, BEL: ICST (Institute for Computer Sciences, Social-Informatics and Telecom...

    2009

  5. [13]

    Traffic data collection under mixed traffic conditions using video image processing,

    C. Mallikarjuna, A. Phanindra, and K. R. Rao, “Traffic data collection under mixed traffic conditions using video image processing,”Journal of Transportation Engineering, vol. 135, no. 4, pp. 174–182, 2009

    2009

  6. [14]

    Ua-detrac: A new benchmark and protocol for multi-object detection and tracking,

    L. Wen, D. Du, Z. Caiet al., “Ua-detrac: A new benchmark and protocol for multi-object detection and tracking,”Computer Vision and Image Understanding, vol. 193, p. 102907, 2020

    2020

  7. [15]

    Automated approach for computer vision-based vehicle movement classification at traffic intersections,

    U. Jana, J. P. D. Karmakar, P. Chakrabortyet al., “Automated approach for computer vision-based vehicle movement classification at traffic intersections,”Future Transportation, vol. 3, no. 2, pp. 708–725, 2023

    2023

  8. [16]

    Optimized video track- ing for automated vehicle turning movement counts,

    F. B ´elisle, N. Saunier, G.-A. Bilodeauet al., “Optimized video track- ing for automated vehicle turning movement counts,”Transportation Research Record, vol. 2645, no. 1, pp. 104–112, 2017

    2017

  9. [17]

    Learning semantic scene models from observing activity in visual surveillance,

    D. Makris and T. Ellis, “Learning semantic scene models from observing activity in visual surveillance,”IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), vol. 35, no. 3, pp. 397–408, 2005

    2005

  10. [18]

    Learning semantic scene models by trajectory analysis,

    X. Wang, K. Tieu, and E. Grimson, “Learning semantic scene models by trajectory analysis,” inComputer Vision – ECCV 2006, A. Leonardis, H. Bischof, and A. Pinz, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2006, pp. 110–123

    2006

  11. [19]

    Trajectory analysis and semantic region modeling using nonparametric hierarchical bayesian models,

    X. Wang, K. T. Ma, G.-W. Nget al., “Trajectory analysis and semantic region modeling using nonparametric hierarchical bayesian models,”Int. J. Comput. Vision, vol. 95, no. 3, p. 287–312, Dec. 2011

    2011

  12. [20]

    A survey of vision-based trajectory learning and analysis for surveillance,

    B. T. Morris and M. M. Trivedi, “A survey of vision-based trajectory learning and analysis for surveillance,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 18, no. 8, pp. 1114–1127, 2008

    2008

  13. [21]

    Trajectory learning for activity understanding: Unsupervised, multilevel, and long-term adaptive approach,

    ——, “Trajectory learning for activity understanding: Unsupervised, multilevel, and long-term adaptive approach,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, no. 11, pp. 2287– 2301, 2011. 14

    2011

  14. [22]

    Looking at intersections: A survey of intersection monitoring, behavior and safety analysis of recent studies,

    M. S. Shirazi and B. T. Morris, “Looking at intersections: A survey of intersection monitoring, behavior and safety analysis of recent studies,” Trans. Intell. Transport. Syst., vol. 18, no. 1, p. 4–24, Jan. 2017

    2017

  15. [23]

    The 8th AI City Challenge,

    S. Wang, D. C. Anastasiu, Z. Tanget al., “The 8th AI City Challenge,” inThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2024

    2024

  16. [24]

    Enhanced vehicle movement counting at intersections via a self-learning fisheye camera system,

    M. Adl, R. Ahmed, C. Vidalet al., “Enhanced vehicle movement counting at intersections via a self-learning fisheye camera system,” IEEE Access, vol. 12, pp. 77 947–77 958, 2024

    2024

  17. [25]

    Towards real-time traffic movement count and trajectory reconstruction using virtual traffic lanes,

    A. Abdelhalim and M. Abbas, “Towards real-time traffic movement count and trajectory reconstruction using virtual traffic lanes,” in2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020, pp. 2527–2533

    2020

  18. [26]

    Intelligent highway adap- tive lane learning system in multiple rois of surveillance camera video,

    M. Qiu, L. Christopher, S. Y .-P. Chienet al., “Intelligent highway adap- tive lane learning system in multiple rois of surveillance camera video,” IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 8, pp. 8591–8601, 2024

    2024

  19. [27]

    J. B. MacQueen,Some methods for classification and analysis of multivariate observations. University of California Press, 1967, pp. 281–297

    1967

  20. [28]

    A density-based algorithm for discovering clusters in large spatial databases with noise,

    M. Ester, H.-P. Kriegel, J. Sanderet al., “A density-based algorithm for discovering clusters in large spatial databases with noise,” inProceedings of the Second International Conference on Knowledge Discovery and Data Mining, ser. KDD’96. AAAI Press, 1996, p. 226–231

    1996

  21. [29]

    McLachlan and D

    G. McLachlan and D. Peel,Finite Mixture Models. Wiley, Sept 2000

    2000

  22. [30]

    Maximum likelihood from incomplete data via the em algorithm,

    A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood from incomplete data via the em algorithm,”Journal of the Royal Statistical Society. Series B (Methodological), vol. 39, no. 1, pp. 1–38, 1977

    1977

  23. [31]

    The quickhull algorithm for convex hulls,

    C. B. Barber, D. P. Dobkin, and H. Huhdanpaa, “The quickhull algorithm for convex hulls,”ACM Trans. Math. Softw., vol. 22, no. 4, p. 469–483, Dec. 1996

    1996

  24. [32]

    Detrs beat yolos on real-time object detection,

    Y . Zhao, W. Lv, S. Xuet al., “Detrs beat yolos on real-time object detection,” in2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 16 965–16 974

    2024

  25. [33]

    Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments,

    G. Varma, A. Subramanian, A. Namboodiriet al., “Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments,” in2019 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019, pp. 1743–1751

    2019

  26. [34]

    The bengaluru mobility challenge, 2024,

    B. T. Police, “The bengaluru mobility challenge, 2024,” 2024

    2024

  27. [35]

    Bot-sort: Robust associa- tions multi-pedestrian tracking,

    N. Aharon, R. Orfaig, and B.-Z. Bobrovsky, “Bot-sort: Robust associa- tions multi-pedestrian tracking,” 2022

    2022

  28. [36]

    Finding a

    V . Satopaa, J. Albrecht, D. Irwinet al., “Finding a ”kneedle” in a haystack: Detecting knee points in system behavior,” in2011 31st International Conference on Distributed Computing Systems Workshops, 2011, pp. 166–171

    2011

  29. [37]

    Estimating the dimension of a model,

    G. Schwarz, “Estimating the dimension of a model,”The Annals of Statistics, vol. 6, no. 2, pp. 461–464, 1978

    1978

  30. [38]

    Image segmentation using deep learning: A survey,

    S. Minaee, Y . Boykov, F. Porikliet al., “Image segmentation using deep learning: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 7, pp. 3523–3542, 2022

    2022

  31. [39]

    An anova test for functional data,

    A. Cuevas, M. Febrero, and R. Fraiman, “An anova test for functional data,”Computational Statistics & Data Analysis, vol. 47, no. 1, pp. 111–122, 2004

    2004

  32. [40]

    Tukey’s honestly significant difference (hsd) test,

    H. Abdi and L. J. Williams, “Tukey’s honestly significant difference (hsd) test,” inEncyclopedia of Research Design, N. J. Salkind, Ed. Thousand Oaks, CA: SAGE Publications, 2010, pp. 1–5

    2010

  33. [41]

    Problems with using the normal distri- bution – and ways to improve quality and efficiency of data analysis,

    E. Limpert and W. A. Stahel, “Problems with using the normal distri- bution – and ways to improve quality and efficiency of data analysis,” PLOS ONE, vol. 6, no. 7, pp. 1–8, 07 2011

    2011

  34. [42]

    A generalization of the wilcoxon signed- rank test and its applications,

    S. M. Taheri and G. Hesamian, “A generalization of the wilcoxon signed- rank test and its applications,”Statistical Papers, vol. 54, no. 2, pp. 457–470, May 2013

    2013

  35. [43]

    Overview of friedman’s test and post-hoc analysis,

    D. G. Pereira, A. Afonso, and F. M. Medeiros, “Overview of friedman’s test and post-hoc analysis,”Communications in Statistics - Simulation and Computation, vol. 44, no. 10, pp. 2636–2653, 2015

    2015

  36. [44]

    Distribution-free multiple comparisons,

    P. B. Nemenyi, “Distribution-free multiple comparisons,” Ph.D. disser- tation, Princeton University, 1963

    1963

  37. [45]

    The geh measure and quality of the highway assignment models,

    O. Feldman, “The geh measure and quality of the highway assignment models,” inProceedings of the European Transport Conference 2012. Glasgow, Scotland, UK: Association for European Transport, October 8–10 2012, tRID Accession Number: 1324899

    2012

  38. [46]

    Texas Department of Transportation,Traffic Safety Program Manual: Chapter 13, Microsimulation Analysis, Section 5: Calibration, Texas Department of Transportation, 2024, accessed: 2026-04-19

    2024

  39. [47]

    Sim ´eoni, H

    O. Sim ´eoni, H. V . V o, M. Seitzeret al., “DINOv3,” 2025

    2025

  40. [48]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.”Journal of machine learning research, vol. 9, no. 11, 2008. 1 Supplementary Material Unsupervised Detection of Entry and Exit Regions from Vehicle Trajectories for Camera-Agnostic Turning Movement Counts S-I. PREP...

    2008

  41. [49]

    Filter Based on Length

  42. [50]

    S1: Three-stage trajectory preprocessing pipeline applied to a representative camera feed (see Step 1 in Fig

    Filter Based on Neighbors Final Trajectories Sub-steps for Trajectory Preprocessing Fig. S1: Three-stage trajectory preprocessing pipeline applied to a representative camera feed (see Step 1 in Fig. 1 of the main text). Each stage progressively removes spurious trajectories. 0...

  43. [51]

    Preprocessing Temporal and displacement filtersO(N) Neighborhood filterO(NlogN)

  44. [52]

    ROI Estimation Coordinate extremaO(N L)

  45. [53]

    Point Selection Extractspoints per endO(N s)

  46. [54]

    Exclusion Zone Point filteringO(N s)

  47. [55]

    Clustering (K-Means)Kiterations over boundary pointsO(KN s)

  48. [56]

    All subsequent steps operate on the2N sselected points, wheresis a small constant (s= 7in the recommended configuration)

    Polygon Construction Convex hull per clusterO(N slogN s) Grid cell assignmentO(n 2) Total (region estimation)O(N L) Classification (per trajectory)Point-in-polygon containmentO(1) Classification (Ntrajectories)O(N) The dominant term in region estimation is the ROI computation,...

    1920

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.