Back

Performance Measures and a Data Set for Multi-target, Multi-camera Tracking

Ergys RistaniFrancesco SoleraRoger S. ZouRita CucchiaraCarlo Tomasi

article2016ECCV3,137 citations

Presents unified identification precision-recall metrics and a large-scale, eight-camera annotated dataset to establish a rigorous benchmark for multi-target, multi-camera tracking.

Listen

The paper addresses shortcomings in how multi-target multi-camera tracking systems are evaluated and tested. Existing measures such as CLEAR MOT and MCTA count discrete events like identity switches or handovers, yet these counts often fail to reflect how well a tracker maintains correct identities over time, which matters most in security, surveillance, and sports applications. At the same time, available benchmark datasets are small, short, or recorded under controlled conditions that do not capture realistic blind spots, lighting variation, and thousands of distinct people.

The work therefore sets out to supply three practical contributions: a pair of identity-based performance measures, a large-scale annotated dataset, and an open reference tracker that can serve as a baseline. The authors first derive identification precision, identification recall, and their harmonic mean IDF1 from an optimal bipartite matching that minimizes the total number of mismatched frames between ground-truth and computed trajectories. They then record and fully annotate more than two million frames from eight synchronized 1080p cameras on the Duke University campus, capturing 2,834 distinct identities across 85 minutes of heavy pedestrian traffic. Finally, they extend an existing single-camera correlation-clustering tracker to the multi-camera setting and run it end-to-end on both the new data and an earlier benchmark.

When evaluated on the new dataset, the identity-based scores reveal that traditional event counts correlate only weakly with actual identity preservation. In roughly three-quarters of the handovers produced by the reference system, short trajectory fragments near camera boundaries cause conventional measures to assign large penalties where the overall identity match is essentially correct, or to assign none where it is largely wrong. On the earlier NLPR benchmark the reference system ranks second among six published methods when scored with the prevailing MCTA metric, even though it uses only simple color histograms rather than learned features. On the Duke data the same system achieves an overall IDF1 of 43 percent, with per-camera scores ranging from 46 percent to 75 percent; these numbers establish a concrete baseline for future work.

The new measures therefore give application-oriented users a clearer signal of how often a tracker correctly reports who is where, while the dataset supplies the volume and realism needed to expose weaknesses that smaller collections hide. System designers can now compare components directly on the same large-scale test rather than on ground-truth single-camera trajectories. The authors recommend that future papers report both the traditional and the identity-based scores on the DukeMTMC test set and release code so that incremental improvements can be measured consistently.

The dataset covers only one outdoor scene under overcast conditions and uses mostly non-overlapping views, so trackers that exploit overlapping fields of view or indoor lighting may need additional test material. Nevertheless, the scale of the annotations and the explicit baseline results give the community a solid foundation for measuring genuine progress in multi-camera identity tracking.

Duke UniversityUniversity of Modena and Reggio Emiliaframesmultitargetreference softwareMulti-Target Multi-Camera TrackingID PrecisionID Recall
arXiv: 1609.01775
  • Paper: Simple online and realtime tracking, Alex Bewley et al. (2016). Reading this foundational tracking-by-detection paper first provides essential background on associating bounding boxes across frames using motion and overlap cues before addressing multi-camera settings.
  • Paper: Simple online and realtime tracking with a deep association metric, Nicolai Wojke et al. (2017). Understanding this extension of SORT with appearance metrics clarifies how identity switches are managed in multi-object tracking prior to evaluating full multi-camera pipelines.
Cover for Performance Measures and a Data Set for Multi-target, Multi-camera Tracking

Abstract

To help accelerate progress in multi-target, multi-camera tracking systems, we present (i) a new pair of precision-recall measures of performance that treats errors of all types uniformly and emphasizes correct identification over sources of error; (ii) the largest fully-annotated and calibrated data set to date with more than 2 million frames of 1080p, 60fps video taken by 8 cameras observing more than 2,700 identities over 85 minutes; and (iii) a reference software system as a comparison baseline. We show that (i) our measures properly account for bottom-line identity match performance in the multi-camera setting; (ii) our data set poses realistic challenges to current trackers; and (iii) the performance of our system is comparable to the state of the art.

Topics

framesmultitargetreference softwaresoftwaretracking performancetracking systemsoftware systems

Keywords

Multi-Target Multi-Camera TrackingID PrecisionID RecallIDF1DukeMTMCMCTACLEAR MOTFragmentationHandover ErrorCalibrated Data Set

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • 3 Performance Measures
  • 3.1 Within-Camera Issues
  • 3.2 Handover Issues
  • 3.3 The Truth-To-Result Match
  • 3.4 Identification Precision, Identification Recall, and F1F_1 Score
  • 3.5 Additional Comparative Remarks
  • 4 Data Set
  • 5 Reference System
  • 6 Experiments
  • 7 Conclusion
  • References

Knowls

  1. Knowl 1 — Optimal Truth-to-Result Bipartite Matching for Tracking Evaluation

    model/method

    To evaluate multi-target multi-camera (MTMC) tracking based on target identity preservation over time rather than local frame-to-frame switching events, ground-truth identities are mapped to computed tracker identities via a global bipartite match that minimizes the total count of mismatched frames over the entire video sequence.

    Let ATAT denote the set of all ground-truth trajectories (true identities) and ACAC denote the set of all computed trajectories (tracker output identities). A bipartite graph G=(VT,VC,E)G = (V_T, V_C, E) is constructed with:

    • Vertex set VTV_T containing one regular node τ\tau for each true trajectory τ∈AT\tau \in AT, and one irregular "false positive" node fγ+f_\gamma^+ for each computed trajectory γ∈AC\gamma \in AC.
    • Vertex set VCV_C containing one regular node γ\gamma for each computed trajectory γ∈AC\gamma \in AC, and one irregular "false negative" node fτ−f_\tau^- for each true trajectory τ∈AT\tau \in AT.
    • Edge set EE containing an edge between every pair of regular nodes (τ,γ)(\tau, \gamma) whose trajectories overlap in time, as well as edges (τ,fτ−)(\tau, f_\tau^-) for each τ∈AT\tau \in AT and (fγ+,γ)(f_\gamma^+, \gamma) for each γ∈AC\gamma \in AC.

    For a frame tt, let τ(t)\tau(t) and γ(t)\gamma(t) represent the spatial positions/bounding boxes of trajectories τ\tau and γ\gamma across their respective frame support sets TτT_\tau and TγT_\gamma. The binary spatial miss indicator m(τ,γ,t,Δ)∈{0,1}m(\tau, \gamma, t, \Delta) \in \{0, 1\} is defined as:

    • For regular pairs (τ,γ)(\tau, \gamma), m(τ,γ,t,Δ)=1m(\tau, \gamma, t, \Delta) = 1 if the spatial overlap fails threshold Δ\Delta. In the image plane, a miss occurs when the intersection-over-union (IoU) of detection bounding boxes is strictly less than Δ∈(0,1)\Delta \in (0, 1). On the reference ground plane, a miss occurs when the Euclidean distance between ground locations exceeds Δ=1 m\Delta = 1\text{ m}. If spatial overlap is sufficient, m(τ,γ,t,Δ)=0m(\tau, \gamma, t, \Delta) = 0.
    • If either node is irregular (fτ−f_\tau^- or fγ+f_\gamma^+), all detections in the other regular trajectory count as misses (m=1m = 1).
    • Matching two irregular nodes incurs zero cost.

    The cost on edge (τ,γ)∈E(\tau, \gamma) \in E is the sum of false negative and false positive detection frames:

    c(τ,γ,Δ)=∑t∈Tτm(τ,γ,t,Δ)+∑t∈Tγm(τ,γ,t,Δ)c(\tau, \gamma, \Delta) = \sum_{t \in T_\tau} m(\tau, \gamma, t, \Delta) + \sum_{t \in T_\gamma} m(\tau, \gamma, t, \Delta)

    Solving minimum-cost bipartite matching on GG yields a bijective assignment:

    • Regular matches (τ,γ)(\tau, \gamma) form the set of True Positive ID matches (IDTPIDTP), defining optimal assignment functions γm(τ)\gamma_m(\tau) and τm(γ)\tau_m(\gamma) between matched subsets MT⊆ATMT \subseteq AT and MC⊆ACMC \subseteq AC.
    • Matches (fγ+,γ)(f_\gamma^+, \gamma) define False Positive IDs (IDFPIDFP).
    • Matches (τ,fτ−)(\tau, f_\tau^-) define False Negative IDs (IDFNIDFN).

  2. Knowl 2 — Identification Precision, Identification Recall, and IDF1 Tracking Metrics

    definition

    Given the optimal truth-to-result bipartite match between ground-truth identities ATAT and computed tracker identities ACAC, tracking performance is evaluated by the cumulative duration over which target identities are correctly maintained.

    Let m(τ,γ,t,Δ)∈{0,1}m(\tau, \gamma, t, \Delta) \in \{0, 1\} be the binary spatial miss indicator at frame tt with threshold Δ\Delta, len(τ)\text{len}(\tau) be the total frame duration of true trajectory τ\tau, and len(γ)\text{len}(\gamma) be the frame duration of computed trajectory γ\gamma. Let γm(τ)\gamma_m(\tau) and τm(γ)\tau_m(\gamma) denote the optimal assignment mappings from the truth-to-result bipartite match.

    The aggregate counts of identification error frames and correctly identified frames are:

    IDFN=∑τ∈AT∑t∈Tτm(τ,γm(τ),t,Δ)IDFN = \sum_{\tau \in AT} \sum_{t \in T_\tau} m(\tau, \gamma_m(\tau), t, \Delta)

    IDFP=∑γ∈AC∑t∈Tγm(τm(γ),γ,t,Δ)IDFP = \sum_{\gamma \in AC} \sum_{t \in T_\gamma} m(\tau_m(\gamma), \gamma, t, \Delta)

    IDTP=∑τ∈ATlen(τ)−IDFN=∑γ∈AClen(γ)−IDFPIDTP = \sum_{\tau \in AT} \text{len}(\tau) - IDFN = \sum_{\gamma \in AC} \text{len}(\gamma) - IDFP

    From these counts, three performance measures are defined:

    • Identification Precision (IDPIDP): The fraction of computed detections that are correctly identified:

    IDP=IDTPIDTP+IDFPIDP = \frac{IDTP}{IDTP + IDFP}

    • Identification Recall (IDRIDR): The fraction of true detections that are correctly identified:

    IDR=IDTPIDTP+IDFNIDR = \frac{IDTP}{IDTP + IDFN}

    • Identification F1F_1 Score (IDF1IDF_1): The harmonic mean of IDPIDP and IDRIDR, measuring correctly identified detections relative to the average number of ground-truth and computed detections:

    IDF1=2⋅IDTP2⋅IDTP+IDFP+IDFNIDF_1 = \frac{2 \cdot IDTP}{2 \cdot IDTP + IDFP + IDFN}

    These metrics provide three core guarantees: bijectivity (one-to-one mapping between true and computed tracks), optimality (evaluator matches identities in the manner most favorable to the tracker), and consistency (all error types are measured in the uniform currency of misassigned frames).

  3. Knowl 3 — The DukeMTMC Multi-Target Multi-Camera Tracking Benchmark Dataset

    experimental setup

    The DukeMTMC benchmark is an outdoor, multi-target multi-camera video dataset collected on the Duke University campus for evaluating MTMC tracking algorithms.

    Key dataset characteristics include:

    • Scale: 8 synchronized static cameras recording at 1080p resolution (1920×10801920 \times 1080) at 60 frames per second (fps) for 85 minutes per camera. The dataset contains over 2 million video frames, more than 10 total hours of video, and over 30 hours of cumulative trajectory time.
    • Identities and Paths: 2,834 unique pedestrian identities and 6,791 single-camera trajectories, averaging 2.5 single-camera trajectories per identity (up to 7). 891 individuals appear in only a single camera view.
    • Camera Topology: Views are largely non-overlapping and disjoint, except for two camera pairs (2–8 and 3–5) that share small overlapping regions traversed by ~100 people. The dataset contains 4,159 total camera hand-overs, with up to 50 people simultaneously traversing blind spots.
    • Crowd Complexity and Occlusions: Pedestrian density reaches up to 54 people per frame in a single view. Over 1,800 self-occlusion events (with ≥50%\ge 50\% bounding-box overlap) occur, with an average duration of 60 frames.
    • Calibration and Annotations: Trajectories are manually annotated at foot contact points across approximately 100,000 key points, interpolated per-frame with bounding boxes and calibrated ground-plane world coordinates via homographies.
    • Data Split: The first 5 minutes across all 8 cameras are allocated for training and validation, while the remaining 80 minutes per camera are reserved for testing.

  4. Knowl 4 — Hierarchical Correlation Clustering Framework for MTMC Tracking

    model/method

    The reference MTMC tracker decomposes multi-camera multi-target data association into a three-layer hierarchical optimization pipeline:

    1. Detections to Tracklets: Pedestrian detections from consecutive frames are linked into short, reliable tracklets based on motion continuity.
    2. Tracklets to Single-Camera Trajectories: Tracklets within the same camera view are linked across occlusion gaps into full single-camera trajectories.
    3. Single-Camera Trajectories to Multi-Camera Identities: Trajectories across different camera views are associated into cross-camera identities.

    At each layer, association is cast as a correlation clustering problem on a graph G=(V,E)G = (V, E). Nodes VV represent observations at that level (detections, tracklets, or trajectories), and edges connect pairs (i,j)(i, j) with affinity weights wij∈[−1,1]∪{−∞,+∞}w_{ij} \in [-1, 1] \cup \{-\infty, +\infty\}. A positive weight represents evidence favoring the same identity, while a negative weight penalizes it, computed from striped color histogram appearance features and space-time travel constraints.

    The graph partitioning is solved as a Binary Integer Program (BIP) that partitions VV to maximize:

    ∑(i,j)∈Ewithinwij−∑(i,j)∈Ebetweenwij\sum_{(i, j) \in E_{\text{within}}} w_{ij} - \sum_{(i, j) \in E_{\text{between}}} w_{ij}

    where EwithinE_{\text{within}} consists of edges connecting nodes assigned to the same identity cluster, and EbetweenE_{\text{between}} connects nodes assigned to different clusters.

    To handle long sequences efficiently despite the NP-hardness of BIP correlation clustering, the optimization runs over sliding temporal windows, propagating prior solutions across overlapping windows as extended observations, and pre-clustering nodes into spatio-temporal and appearance subgroups.

  5. Knowl 5 — Handover Difficulty Quantification via Single- vs. Multi-Camera Error Differential

    equation

    To isolate the performance cost incurred by associating target identities across different camera views compared to tracking within individual cameras alone, tracking errors can be evaluated independently at single-camera and multi-camera levels.

    Let EME_M be the total identification error frames when the truth-to-result match is computed globally across the entire multi-camera network:

    EM=IDFPM+IDFNME_M = IDF P_M + IDF N_M

    Let ESE_S be the sum of identification error frames obtained when the truth-to-result match is computed independently for each camera and then summed over all cameras:

    ES=IDFPS+IDFNSE_S = IDF P_S + IDF N_S

    Because enforcing cross-camera identity consistency introduces additional constraints over independent per-camera matching, the error differential is guaranteed to be nonnegative:

    EM−ES≥0E_M - E_S \ge 0

    Similarly, single-camera precision, recall, and F1F_1 scores upper-bound their multi-camera counterparts:

    IDPS−IDPM≥0,IDRS−IDRM≥0,F1S−F1M≥0IDP_S - IDP_M \ge 0, \quad IDR_S - IDR_M \ge 0, \quad F_{1S} - F_{1M} \ge 0

    These non-negative gaps isolate and quantify the specific difficulty of target handover across cameras.

  6. Knowl 6 — Performance of Reference MTMC Tracker on the NLPR MCT Benchmark

    data/table

    The reference MTMC tracking baseline was compared against existing multi-camera trackers on the NLPR MCT dataset across four subsets (NLPR 1 to NLPR 4) using the Multi-Camera Object Tracking Accuracy (MCTA) metric. In this standard evaluation setup, all algorithms received the same ground-truth single-camera trajectories as input.

    Systems NLPR 1 NLPR 2 NLPR 3 NLPR 4 Avg. Rank
    USC 0.9152 0.9132 0.5163 0.7052 2.25
    Reference Baseline (Ours) 0.7967 0.7336 0.6543 0.7616 2.50
    GE 0.8353 0.7034 0.7417 0.3845 2.75
    hfutdspmct 0.7425 0.6544 0.7368 0.3945 3.50
    CRIPAC-MCT 0.6617 0.5907 0.7105 0.5703 4.00
    Adb-Team 0.3204 0.3456 0.1382 0.1563 6.00

    The reference correlation-clustering baseline ranks second overall with an average rank of 2.50 across the four benchmarks using default appearance features (striped color histograms), ranking behind USC (2.25 average rank), which incorporates discriminatively learned appearance models.

  7. Knowl 7 — Single-Camera and Multi-Camera Tracking Baseline Performance on DukeMTMC

    data/table

    The reference MTMC tracker was evaluated on the DukeMTMC dataset across all 8 individual camera views (single-camera setting) and on the complete multi-camera configuration. The reported scores include standard CLEAR MOT measures (FP%FP\%, FN%FN\%, Identity Switches IDSIDS, Fragmentations FRGFRG, MOTA%MOTA\%, MOTP%MOTP\%, Ground Truth tracks GTGT, Mostly Tracked MTMT, Mostly Lost MLML) alongside identity-based metrics (IDP%IDP\%, IDR%IDR\%, IDF1%IDF_1\%).

    Cam FP% FN% IDS FRG MOTA% MOTP% GT MT ML IDP% IDR% IDF1%
    1 9.70 52.90 178 366 37.36 67.57 1175 105 128 79.17 44.97 57.36
    2 21.48 29.19 866 1929 49.17 61.70 1106 416 50 69.11 63.78 66.34
    3 7.04 39.39 134 336 53.50 63.57 501 229 42 81.46 55.11 65.74
    4 10.61 33.42 107 403 55.92 66.51 390 128 21 79.23 61.16 69.03
    5 3.48 23.38 162 292 73.09 70.52 644 396 33 84.86 67.97 75.48
    6 38.62 48.21 1426 3370 12.94 48.62 1043 207 91 48.35 43.71 45.91
    7 8.28 29.57 296 675 62.03 60.73 678 373 53 85.23 67.08 75.07
    8 1.29 61.69 270 365 36.98 69.07 1254 369 236 90.54 35.86 51.37
    1–8 (Upper bound) – – – – – – – – – 72.25 50.96 59.77
    1–8 (Baseline) – – – – – – – – – 52.35 36.46 42.98

    The multi-camera upper bound reflects single-camera tracking evaluated without enforcing cross-camera ID consistency (IDF1=59.77%IDF_1 = 59.77\%), while the fully integrated multi-camera baseline achieves IDF1=42.98%IDF_1 = 42.98\% (IDP=52.35%IDP = 52.35\%, IDR=36.46%IDR = 36.46\%). Single-camera difficulty rankings differ between MOTAMOTA and IDF1IDF_1 (e.g., cameras 2 and 3), reflecting differences between detection/event counting and persistent identity matching.

  8. Knowl 8 — High Frequency of Spurious Handover Penalties in Boundary-Based Tracking Evaluation

    empirical result

    Traditional event-based multi-camera tracking measures evaluate handover errors based strictly on identity labels at the transition boundary (the last frame of camera I and the first frame of camera II). This makes them brittle to brief, isolated identity fluctuations near camera transitions.

    In an empirical evaluation of 5,549 handovers produced by the baseline tracker on the DukeMTMC dataset, 74.2% (4,119 out of 5,549 handovers) suffer from evaluation discrepancies caused by boundary artifacts:

    • False Handover Fragmentations (43.8%): In 2,430 handovers (1,692 cases / 30.5% and 738 cases / 13.3%), target trajectories were handed over correctly between cameras for nearly their entire duration, but a brief erroneous fragment near the boundary triggered a severe handover fragmentation penalty under event-based metrics.
    • False Handover Merges (28.1%): In 1,561 handovers (1,065 cases / 19.2% and 496 cases / 8.9%), a handover merge error was charged due to a short boundary fragment despite the overall handover being correct.
    • Missed Handover Errors (2.3%): In 128 handovers (70 cases / 1.3% and 58 cases / 1.0%), event-based measures failed to penalize an incorrect handover because a single spurious boundary fragment happened to match labels across cameras.

    Truth-to-result bipartite matching resolves these distortions by penalizing identity errors proportionally to their temporal duration rather than their occurrence at boundaries.

  9. Knowl 9 — Weak Correlation Between Trajectory Event Errors and Identity Precision/Recall

    empirical result

    An empirical evaluation on the DukeMTMC dataset demonstrates that traditional event-based error counts (trajectory fragmentations and merges) correlate poorly with identity-based tracking metrics (IDPIDP and IDRIDR):

    • Correlation between trajectory fragmentations and IDRIDR: −0.24-0.24
    • Correlation between trajectory merges and IDRIDR: −0.05-0.05
    • Correlation between trajectory fragmentations and IDPIDP: −0.38-0.38
    • Correlation between trajectory merges and IDPIDP: −0.41-0.41

    These low correlation values indicate that counting the number of local identity switch events provides an inadequate proxy for measuring how long target identities are correctly preserved over time.

  10. Knowl 10 — Limitations of the DukeMTMC Benchmark Dataset

    limitation

    The DukeMTMC benchmark dataset has three specific domain limitations:

    1. Single Environment and Lighting: All footage is recorded outdoors on a single university campus under diffuse, overcast illumination, which reduces extreme lighting variations (such as strong shadows or direct sunlight) that could further complicate tracking.
    2. Static Camera Network: Video is recorded exclusively using fixed, stationary cameras, excluding mobile cameras, wearable sensors, or pan-tilt-zoom (PTZ) camera systems.
    3. Mostly Disjoint Fields of View: Only two camera pairs have small overlapping spatial regions, while the remaining camera views are disjoint. Consequently, the dataset does not evaluate methods designed specifically for dense multi-view geometric fusion across overlapping fields of view.

Coverage note — No substantial contributed material was omitted; all primary contributions—including the truth-to-result matching formulation, ID precision/recall/IDF1 metrics, DukeMTMC dataset specifications and baseline results, the reference hierarchical tracking system, handover error pathology analysis, and dataset limitations—are fully captured.

References

  1. 1.Ristani, E., Tomasi, C.: Tracking multiple people online and in real time. In: ACCV-12th Asian Conference on Computer Vision, Springer (2014)
  2. 2.Cao, L., Chen, W., Chen, X., Zheng, S., Huang, K.: An equalised global graphical model-based approach for multi-camera object tracking. ArXiv:11502.03532 [cs] (February 2015)
  3. 3.Bernardin, K., Stiefelhagen, R.: Evaluating multiple object tracking performance: the CLEAR MOT metrics. EURASIP Journal on Image and Video Processing (246309) (2008) 1–10
  4. 4.Wu, B., Nevatia, R.: Tracking of multiple, partially occluded humans based on static body part detection. In: Computer Vision and Pattern Recognition, 2006 IEEE Computer Society Conference on. Volume 1., IEEE (2006) 951–958
  5. 5.Milan, A., Schindler, K., Roth, S.: Challenges of ground truth evaluation of multi-target tracking. In: Computer Vision and Pattern Recognition Workshops (CVPRW), 2013 IEEE Conference on, IEEE (2013) 735–742
  6. 6.Kuo, C.H., Huang, C., Nevatia, R.: Inter-camera Association of Multi-target Tracks by On-Line Learned Appearance Affinity Models. In Daniilidis, K., Maragos, P., Paragios, N., eds.: Computer Vision ECCV 2010. Number 6311 in Lecture Notes in Computer Science. Springer Berlin Heidelberg (2010) 383–396
  7. 7.Multi-Camera Object Tracking Challenge: ECCV workshop on visual surveillance and re-identification. http://mct.idealtest.org (2014)
  8. 8.Fleuret, F., Berclaz, J., Lengagne, R., Fua, P.: Multi-Camera People Tracking with a Probabilistic Occupancy Map. IEEE Transactions on Pattern Analysis and Machine Intelligence 30(2) (February 2008) 267–282
  9. 9.Berclaz, J., Fleuret, F., T¨uretken, E., Fua, P.: Multiple object tracking using k-shortest paths optimization. IEEE Transactions on Pattern Analysis and Machine Intelligence (2011)
  10. 10.Ferryman, J., Shahrokni, A.: An overview of the PETS 2009 challenge. (2009)
  11. 11.D’Orazio, T., Leo, M., Mosca, N., Spagnolo, P., Mazzeo, P.L.: A semi-automatic system for ground truth generation of soccer video sequences. In: Advanced Video and Signal Based Surveillance, 2009. AVSS’09. Sixth IEEE International Conference on, IEEE (2009) 559–564
  12. 12.De Vleeschouwer, C., Chen, F., Delannay, D., Parisot, C., Chaudy, C., Martrou, E., Cavallaro, A., et al.: Distributed video acquisition and annotation for sport-event summarization. In: NEM summit 2008:: Towards Future Media Internet. (2008)
  13. 13.Zhang, S., Staudt, E., Faltemier, T., Roy-Chowdhury, A.: A Camera Network Tracking (CamNeT) Dataset and Performance Baseline. In: 2015 IEEE Winter Conference on Applications of Computer Vision (WACV). (January 2015) 365–372
  14. 14.Per, J., Kenk, V.S., Mandeljc, R., Kristan, M., Kovaˇciˇc, S.: Dana36: A multi-camera image dataset for object identification in surveillance scenarios. In: Advanced Video and Signal-Based Surveillance (AVSS), 2012 IEEE Ninth International Conference on, IEEE (2012) 64–69
  15. 15.Benenson, R., Omran, M., Hosang, J., Schiele, B.: Ten years of pedestrian detection, what have we learned? In: ECCV 2014 Workshops. Volume 8926. (2015) 613–627
  16. 16.Leal-Taix´e, L., Milan, A., Reid, I., Roth, S., Schindler, K.: Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv:1504.01942 [cs] (April 2015) arXiv: 1504.01942.
  17. 17.Bredereck, M., Jiang, X., Korner, M., Denzler, J.: Data association for multi-object Tracking-by-Detection in multi-camera networks. In: 2012 Sixth International Conference on Distributed Smart Cameras (ICDSC). (October 2012) 1–6
  18. 18.Cai, Y., Medioni, G.: Exploring context information for inter-camera multiple target tracking. In: 2014 IEEE Winter Conference on Applications of Computer Vision (WACV). (March 2014) 761–768
  19. 19.Chen, K.W., Lai, C.C., Lee, P.J., Chen, C.S., Hung, Y.P.: Adaptive Learning for Target Tracking and True Linking Discovering Across Multiple Non-Overlapping Cameras. IEEE Transactions on Multimedia 13(4) (August 2011) 625–638
  20. 20.Chen, X., An, L., Bhanu, B.: Multitarget Tracking in Nonoverlapping Cameras Using a Reference Set. IEEE Sensors Journal 15(5) (May 2015) 2692–2704
  21. 21.Chen, X., Huang, K., Tan, T.: Direction-based stochastic matching for pedestrian recognition in non-overlapping cameras. In: 2011 18th IEEE International Conference on Image Processing (ICIP). (September 2011) 2065–2068
  22. 22.Daliyot, S., Netanyahu, N.S.: A Framework for Inter-camera Association of Multi-target Trajectories by Invariant Target Models. In Park, J.I., Kim, J., eds.: Computer Vision - ACCV 2012 Workshops. Number 7729 in Lecture Notes in Computer Science. Springer Berlin Heidelberg (2013) 372–386
  23. 23.Das, A., Chakraborty, A., Roy-Chowdhury, A.K.: Consistent re-identification in a camera network. In: Computer Vision–ECCV 2014. Springer (2014) 330–345
  24. 24.Gilbert, A., Bowden, R.: Tracking Objects Across Cameras by Incrementally Learning Inter-camera Colour Calibration and Patterns of Activity. In Leonardis, A., Bischof, H., Pinz, A., eds.: Computer Vision ECCV 2006. Number 3952 in Lecture Notes in Computer Science. Springer Berlin Heidelberg (2006) 125–136
  25. 25.Javed, O., Shafique, K., Rasheed, Z., Shah, M.: Modeling inter-camera spacetime and appearance relationships for tracking across non-overlapping views. Computer Vision and Image Understanding 109(2) (February 2008) 146–162
  26. 26.Makris, D., Ellis, T., Black, J.: Bridging the gaps between cameras. In: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. Volume 2. (June 2004)
  27. 27.Jiuqing, W., Li, L.: Distributed optimization for global data association in non-overlapping camera networks. In: 2013 Seventh International Conference on Distributed Smart Cameras (ICDSC). (October 2013) 1–7
  28. 28.Calderara, S., Cucchiara, R., Prati, A.: Bayesian-competitive consistent labeling for people surveillance. Pattern Analysis and Machine Intelligence, IEEE Transactions on 30(2) (Feb 2008) 354–360
  29. 29.Zhang, S., Zhu, Y., Roy-Chowdhury, A.: Tracking multiple interacting targets in a camera network. Computer Vision and Image Understanding 134 (May 2015) 64–73
  30. 30.Ayazoglu, M., Li, B., Dicle, C., Sznaier, M., Camps, O.: Dynamic subspace-based coordinated multicamera tracking. In: 2011 IEEE International Conference on Computer Vision (ICCV). (November 2011) 2462–2469
  31. 31.Kamal, A., Farrell, J., Roy-Chowdhury, A.: Information Consensus for Distributed Multi-target Tracking. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). (June 2013) 2403–2410
  32. 32.Hamid, R., Kumar, R., Grundmann, M., Kim, K., Essa, I., Hodgins, J.: Player localization using multiple static cameras for sports visualization. In: 2010 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). (June 2010) 731–738
  33. 33.Martinel, N., Micheloni, C., Foresti, G.L.: Saliency weighted features for person re-identification. In: Computer Vision-ECCV 2014 Workshops. Springer International Publishing (2014) 191–208
  34. 34.Zhao, R., Ouyang, W., Wang, X.: Unsupervised salience learning for person re-identification. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). (2013)
  35. 35.Bedagkar-Gala, A., Shah, S.: Multiple person re-identification using part based spatio-temporal color appearance model. In: Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on. (Nov 2011) 1721–1728
  36. 36.Bedagkar-Gala, A., Shah, S.K.: Part-based spatio-temporal model for multi-person re-identification. Pattern Recognition Letters 33(14) (2012) 1908 – 1915 Novel Pattern Recognition-Based Methods for Re-identification in Biometric Context.
  37. 37.Cheng, D., Cristani, M., Stoppa, M., Bazzani, L., Murino, V.: Custom pictorial structures for re-identification. In: Proceedings of the British Machine Vision Conference, BMVA Press (2011) 68.1–68.11 http://dx.doi.org/10.5244/C.25.68.
  38. 38.Baltieri, D., Vezzani, R., Cucchiara, R.: Learning articulated body models for people re-identification. In: Proceedings of the 21st ACM International Conference on Multimedia. MM ’13, New York, NY, USA, ACM (2013) 557–560
  39. 39.Cheng, D., Cristani, M.: Person re-identification by articulated appearance matching. In Gong, S., Cristani, M., Yan, S., Loy, C.C., eds.: Person Re-Identification. Advances in Computer Vision and Pattern Recognition. Springer London (2014) 139–160
  40. 40.Baltieri, D., Vezzani, R., Cucchiara, R.: Mapping appearance descriptors on 3d body models for people re-identification. International Journal of Computer Vision 111(3) (2015) 345–364
  41. 41.Brendel, W., Amer, M., Todorovic, S.: Multiobject tracking as maximum weight independent set. In: Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, IEEE (2011) 1273–1280
  42. 42.Shu, G., Dehghan, A., Oreifej, O., Hand, E., Shah, M.: Part-based multiple-person tracking with partial occlusion handling. In: Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, IEEE (2012) 1815–1821
  43. 43.Wu, B., Nevatia, R.: Detection and tracking of multiple, partially occluded humans by bayesian combination of edgelet based part detectors. International Journal of Computer Vision 75(2) (2007) 247–266
  44. 44.Izadinia, H., Saleemi, I., Li, W., Shah, M.: Mp2t: Multiple people multiple parts tracker. In Fitzgibbon, A.W., Lazebnik, S., Perona, P., Sato, Y., Schmid, C., eds.: ECCV (6). Volume 7577 of Lecture Notes in Computer Science., Springer (2012) 100–114
  45. 45.Pirsiavash, H., Ramanan, D., Fowlkes, C.C.: Globally-optimal greedy algorithms for tracking a variable number of objects. In: Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, IEEE (2011) 1201–1208
  46. 46.Zhang, L., Li, Y., Nevatia, R.: Global data association for multi-object tracking using network flows. In: Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on, IEEE (2008) 1–8
  47. 47.Butt, A.A., Collins, R.T.: Multiple target tracking using frame triplets. In: Computer Vision–ACCV 2012. Springer (2013) 163–176
  48. 48.Chari, V., Lacoste-Julien, S., Laptev, I., Sivic, J.: On pairwise costs for network flow multi-object tracking. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. (2015) 5537–5545
  49. 49.Collins, R.T.: Multitarget data association with higher-order motion models. In: Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, IEEE (2012) 1744–1751
  50. 50.Dehghan, A., Assari, S.M., Shah, M.: Gmmcp tracker: Globally optimal generalized maximum multi clique problem for multiple object tracking. In: CVPR. Volume 1. (2015) 2
  51. 51.Kumar, R., Charpiat, G., Thonnat, M.: Multiple object tracking by efficient graph partitioning. In: Computer Vision–ACCV 2014. Springer (2014) 445–460
  52. 52.Shafique, K., Shah, M.: A noniterative greedy algorithm for multiframe point correspondence. Pattern Analysis and Machine Intelligence, IEEE Transactions on 27(1) (2005) 51–65
  53. 53.Tang, S., Andres, B., Andriluka, M., Schiele, B.: Subgraph decomposition for multi-target tracking. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. (2015) 5033–5041
  54. 54.Wen, L., Li, W., Yan, J., Lei, Z., Yi, D., Li, S.Z.: Multiple target tracking based on undirected hierarchical relation hypergraph. In: Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on, IEEE (2014) 1282–1289
  55. 55.Zamir, A., Dehghan, A., Shah, M.: Gmcp-tracker: Global multi-object tracking using generalized minimum clique graphs. In: Proceedings of the European Conference on Computer Vision (ECCV). (2012)
  56. 56.Yu, S.I., Meng, D., Zuo, W., Hauptmann, A.: The solution path algorithm for identity-aware multi-object tracking. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. (2016) 3871–3879
  57. 57.Yao, J., Odobez, J.M.: Multi-layer background subtraction based on color and texture. In: Computer Vision and Pattern Recognition, 2007. CVPR’07. IEEE Conference on, IEEE (2007) 1–8
  58. 58.Felzenszwalb, P., Girshick, R., McAllester, D., Ramanan, D.: Object detection with discriminatively trained part-based models. Pattern Analysis and Machine Intelligence, IEEE Transactions on 32(9) (Sept 2010) 1627–1645
  59. 59.Bansal, N., Blum, A., Chawla, S.: Correlation clustering. In: Foundations of Computer Science. (2002)
  60. 60.Tan, J.: A note on the inapproximability of correlation clustering. (2008)
  61. 61.Liu, C., Gong, S., Loy, C.C., Lin, X.: Person re-identification: What features are important? In: Computer Vision–ECCV 2012. Workshops and Demonstrations, Springer (2012) 391–401
  62. 62.Chen, W., Cao, L., Chen, X., Huang, K.: A novel solution for multi-camera object tracking. In: Image Processing (ICIP), 2014 IEEE International Conference on, IEEE (2014) 2329–2333

Citation

MLA
Ristani, E., et al. “Performance Measures and a Data Set for Multi-target, Multi-camera Tracking”. Lecture Notes in Computer Science, Springer International Publishing, 2016, pp. 17–35, https://doi.org/10.1007/978-3-319-48881-3_2.
APA
Ristani, E., Solera, F., Zou, R., Cucchiara, R., & Tomasi, C. (2016). Performance Measures and a Data Set for Multi-target, Multi-camera Tracking. In Lecture Notes in Computer Science (pp. 17–35). Springer International Publishing. https://doi.org/10.1007/978-3-319-48881-3_2
Chicago
Ristani, E., F. Solera, R. Zou, R. Cucchiara, and C. Tomasi. 2016. “Performance Measures and a Data Set for Multi-target, Multi-camera Tracking”. In Lecture Notes in Computer Science. Springer International Publishing. https://doi.org/10.1007/978-3-319-48881-3_2.
Harvard
Ristani, E. et al. (2016) “Performance Measures and a Data Set for Multi-target, Multi-camera Tracking”, Lecture Notes in Computer Science. Springer International Publishing, pp. 17–35. Available at: https://doi.org/10.1007/978-3-319-48881-3_2.
Vancouver
1. Ristani E, Solera F, Zou R, Cucchiara R, Tomasi C (2016) Performance Measures and a Data Set for Multi-target, Multi-camera Tracking. In: Lecture Notes in Computer Science. Springer International Publishing, pp 17–35

BibTeX

@inbook{Ristani_2016, title={Performance Measures and a Data Set for Multi-target, Multi-camera Tracking}, ISBN={9783319488813}, ISSN={1611-3349}, url={http://dx.doi.org/10.1007/978-3-319-48881-3_2}, DOI={10.1007/978-3-319-48881-3_2}, booktitle={Computer Vision – ECCV 2016 Workshops}, publisher={Springer International Publishing}, author={Ristani, Ergys and Solera, Francesco and Zou, Roger and Cucchiara, Rita and Tomasi, Carlo}, year={2016}, pages={17–35} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF