pith. sign in

arxiv: 1907.09915 · v1 · pith:IBOIQA3Mnew · submitted 2019-07-16 · 💻 cs.LG · cs.AI· stat.ML

DeepDA: LSTM-based Deep Data Association Network for Multi-Targets Tracking in Clutter

Huajun Liu , Hui Zhang , Christoph Mertz This is my paper

Pith reviewed 2026-05-24 20:42 UTC · model grok-4.3

classification 💻 cs.LG cs.AIstat.ML
keywords data associationmulti-target trackingLSTMdeep learningclutterradar measurementsassociation probability
0
0 comments X

The pith

LSTM network learns to output association probabilities directly from radar measurements and tracks.

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

The paper proposes DeepDA, an LSTM-based deep neural network for the data association step in multi-target tracking. It formally redefines the problem to cover cases with unknown target counts, missed detections, and clutter, then trains the network on simulated radar data to map measurements and existing tracks straight to association probabilities. Training uses backpropagation through time and RMSprop; once trained, the network bypasses explicit models of clutter density or covariance. On simulated scenarios with crossing targets in dense clutter, it reports higher association ratios, fewer identity switches, and lower computation time than classical methods.

Core claim

An LSTM-based deep neural network, after supervised training through BPTT and RMSprop, can compute measurement-to-track association probabilities directly from noisy radar measurements and existing tracks, addressing the combinatorial data association problem for multi-target tracking in clutter.

What carries the argument

LSTM network that learns the mapping from radar measurements and tracks to association probabilities.

If this is right

  • Association can be performed without supplying explicit prior values for clutter density, filter covariance, or gating thresholds.
  • The approach maintains performance when targets cross paths inside complicated clutter.
  • Computation time decreases relative to solving the NP-hard combinatorial problem at each step.
  • Target identity switches are reduced on the simulated test cases.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The learned mapping might be inserted as a drop-in replacement inside existing multi-target trackers that currently use hand-designed association logic.
  • Performance on simulations leaves open whether the same network would require retraining when sensor characteristics change.
  • End-to-end training that includes both association and state estimation could be tested next.

Load-bearing premise

The simulated radar measurements and clutter used for training and testing are representative of real sensor data and environments.

What would settle it

Running the trained network on real radar recordings that contain different clutter statistics or sensor noise than the simulations and measuring a drop in association performance.

Figures

Figures reproduced from arXiv: 1907.09915 by Christoph Mertz, Huajun Liu, Hui Zhang.

Figure 1
Figure 1. Figure 1: shows a typical scenario describing the measure￾ments by a radar where there are false alarms and miss detections. In the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Data association between measurement set and target set We define the pairwise-distance matrix ܥ∋ℝே×ெ to calculate the distance of all measurements and predicted state of all targets. And for any measurement ݖ௜ೖ ௞ ∈ܼ(݇), and any target ݆ , ߬௝ೖ ௞ ∈ ߁)݇ (, the Euclidean distance between the measurement ݅ and the predicted state of target ݆ is C௜௝ ,ܥ∋ ௞ which is defined as ௝௜ܥ ௞ = ฮݖ௜ೖ ௞ −߬௝ೖ ௞ ฮ ଶ (4) For ea… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our proposed data association framework The main idea is to exploit the LSTM’s temporal step-by￾step functionality to predict the assignment for each measurement to one target at a time. The input at each step ݅, next to the hidden state ℎ௜ and the cell state ܿ௜, are the entire feature vector. The input of the network is a pairwise-distance matrix which is an Euclidean distance matrix between t… view at source ↗
Figure 4
Figure 4. Figure 4: Diagram of data association model based on LSTMs Predicting the probability of association between each target and all measurements in one time is shown in [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: LSTM-based DeepDA training framework The dataflow of the DeepDA algorithm’s training is as follows: Algorithm: Training algorithm of DeepDA Network Input: the ground truth of a target at a certain moment, and its corresponding next-time observation; Output: the association probability of this target with current measurements. Step 1: Set the data parameters, select and normalize the data by min-max method … view at source ↗
Figure 7
Figure 7. Figure 7: Simulation of real trajectories and measurements The training data is generated through above simulation environment. Specifically, in a sequence at time k, the training data will include: the input state of the model, all the measurements ܼ௞ and its predicted state ܺ௞ at time k, and the real associated probability ߚ ௞, if the measurement is originated from the target, ߚ= ௞1, else ߚ= ௞0. B. Performance Met… view at source ↗
Figure 9
Figure 9. Figure 9: Multi-target data association results when ܲௗ = 0.8, ܧఒ = 20 [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Loss convergence curve when training The DeepDA network is trained with an LSTM network with 1 hidden layer. The network is trained on the computer configured with NVIDIA Titan Xp GPU and the whole training takes approximately 5 minutes. The loss function is converged during training shown in [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: and [PITH_FULL_IMAGE:figures/full_fig_p006_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The OSPA distance of data association algorithms at different detection probability Similarly, it can be seen from the scale of y-axis in [PITH_FULL_IMAGE:figures/full_fig_p007_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The OSPA distance of different data association algorithms Moreover, it can be concluded from [PITH_FULL_IMAGE:figures/full_fig_p007_12.png] view at source ↗
read the original abstract

The Long Short-Term Memory (LSTM) neural network based data association algorithm named as DeepDA for multi-target tracking in clutters is proposed to deal with the NP-hard combinatorial optimization problem in this paper. Different from the classical data association methods involving complex models and accurate prior knowledge on clutter density, filter covariance or associated gating etc, data-driven deep learning methods have been extensively researched for this topic. Firstly, data association mathematical problem for multitarget tracking on unknown target number, missed detection and clutter, which is beyond one-to-one mapping between observations and targets is redefined formally. Subsequently, an LSTM network is designed to learn the measurement-to-track association probability from radar noisy measurements and exist tracks. Moreover, an LSTM-based data-driven deep neural network after a supervised training through the BPTT and RMSprop optimization method can get the association probability directly. Experimental results on simulated data show a significant performance on association ratio, target ID switching and time-consuming for tracking multiple targets even they are crossing each other in the complicated clutter environment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes DeepDA, an LSTM-based deep neural network for data association in multi-target tracking amid clutter. It formally redefines the association problem for unknown target counts, missed detections, and clutter beyond one-to-one mappings. An LSTM is trained supervised via BPTT and RMSprop to output association probabilities directly from noisy radar measurements and existing tracks. Experiments on simulated data claim superior association ratios, fewer target ID switches, and reduced runtime versus classical methods, even for crossing targets in dense clutter.

Significance. If the experimental claims hold after adding missing details, the work could demonstrate a viable data-driven route to the NP-hard association problem that avoids explicit priors on clutter density or gating. The LSTM's temporal modeling is a natural fit for tracking sequences. Credit is due for attempting a fully supervised, end-to-end formulation on simulated radar scenarios. However, the absence of any baseline comparisons, architecture/loss details, or real-sensor validation substantially reduces the result's immediate significance for the tracking community.

major comments (3)
  1. [Abstract / Experiments] Abstract and Experiments section: the central performance claims (improved association ratio, ID switching, and runtime) are presented without any numerical tables, figures, baseline comparisons (e.g., JPDA, MHT, or nearest-neighbor), or error bars on the same simulated scenarios, rendering the superiority assertion uninspectable and load-bearing for the paper's contribution.
  2. [Method] Method section: the supervised training procedure is described only at the level of 'BPTT and RMSprop'; the loss function, network architecture (layers, hidden size, input encoding of measurements/tracks), and output probability normalization are not specified, which prevents assessment or reproduction of how association probabilities are learned.
  3. [Experiments] Experiments section: all reported results use only simulated radar measurements and clutter; no quantitative matching of clutter density, noise statistics, or missed-detection rates to real sensor data is provided, which directly affects the claim that the method works 'in the complicated clutter environment.'
minor comments (2)
  1. [Abstract] Abstract: the phrase 'significant performance on ... time-consuming' is unclear; it should explicitly state whether this refers to wall-clock runtime, complexity scaling, or another metric.
  2. [Introduction / Problem Statement] The redefinition of the data-association problem is announced but lacks an accompanying equation or formal statement in the abstract; a concise mathematical formulation would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below with point-by-point responses and indicate revisions to the manuscript.

read point-by-point responses

  1. Referee: [Abstract / Experiments] Abstract and Experiments section: the central performance claims (improved association ratio, ID switching, and runtime) are presented without any numerical tables, figures, baseline comparisons (e.g., JPDA, MHT, or nearest-neighbor), or error bars on the same simulated scenarios, rendering the superiority assertion uninspectable and load-bearing for the paper's contribution.

    Authors: We agree that explicit numerical results and baselines are needed for inspectability. The revised manuscript adds Table 1 reporting association ratios, ID switches, and runtimes with error bars from 10 independent runs on identical simulated scenarios, including direct comparisons to JPDA, MHT, and nearest-neighbor under the same clutter densities and crossing-target cases. revision: yes

  2. Referee: [Method] Method section: the supervised training procedure is described only at the level of 'BPTT and RMSprop'; the loss function, network architecture (layers, hidden size, input encoding of measurements/tracks), and output probability normalization are not specified, which prevents assessment or reproduction of how association probabilities are learned.

    Authors: We acknowledge the method section omitted key implementation details. The revision expands Section 3 to specify the cross-entropy loss, a two-layer LSTM with 128 hidden units, input encoding as concatenated normalized measurement and track state vectors, and softmax normalization over association probabilities, enabling full reproduction. revision: yes

  3. Referee: [Experiments] Experiments section: all reported results use only simulated radar measurements and clutter; no quantitative matching of clutter density, noise statistics, or missed-detection rates to real sensor data is provided, which directly affects the claim that the method works 'in the complicated clutter environment.'

    Authors: The work is framed around controlled simulated radar scenarios to isolate the data-association task. We have added a paragraph in the experiments section quantifying how the simulated clutter density (Poisson rate 0.05 per resolution cell), measurement noise (Gaussian sigma 50 m), and missed-detection probability (0.1) align with typical X-band radar parameters reported in the tracking literature, while noting that real-sensor validation remains future work. revision: partial

Circularity Check

1 steps flagged

Association probabilities and performance metrics are outputs of an LSTM fitted via supervised training on the same simulated data distribution used for evaluation

specific steps
  1. fitted input called prediction [Abstract]
    "Moreover, an LSTM-based data-driven deep neural network after a supervised training through the BPTT and RMSprop optimization method can get the association probability directly. Experimental results on simulated data show a significant performance on association ratio, target ID switching and time-consuming for tracking multiple targets even they are crossing each other in the complicated clutter environment."

    The association probabilities are produced by the LSTM whose parameters are fitted via supervised training (BPTT/RMSprop) on simulated measurements and clutter. The reported performance metrics are therefore the direct outputs of this fitted model evaluated on test data from the same simulated distribution, rendering the 'prediction' of good performance statistically forced by construction rather than independently validated.

full rationale

The paper's central method trains an LSTM on simulated radar measurements and clutter to directly output association probabilities, then reports superior association ratio, ID switching, and runtime on (presumably held-out) simulated data. This matches the fitted-input-called-prediction pattern: the claimed predictions reduce to the behavior of a model whose weights were optimized on data drawn from the identical distribution, with no external real-sensor benchmarks or comparisons that would make the results independent of the training inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no equations, parameter counts, or modeling assumptions; free_parameters, axioms, and invented_entities cannot be extracted.

pith-pipeline@v0.9.0 · 5715 in / 1025 out tokens · 20861 ms · 2026-05-24T20:42:44.210844+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · 1 internal anchor

  1. [1]

    Rong Li and Thiagalingam Kirubarajan, Estimation with Applications to Tracking and Navigation: Theory, Algorithms and Software, New York: Wiley, 2002

    Yaakov Bar-Shalom, X. Rong Li and Thiagalingam Kirubarajan, Estimation with Applications to Tracking and Navigation: Theory, Algorithms and Software, New York: Wiley, 2002

    work page 2002

  2. [2]

    Martin Liggins II, David Hall and James Llinas, Handbook of Multisensor Data Fusion, 2nd Edition, CRC Press, 2008

    work page 2008

  3. [3]

    Moving object detection with laser scanners, Journal of Field Robotics,

    Christoph Mertz et al, “Moving object detection with laser scanners, Journal of Field Robotics,” Vol. 30, No. 1, pp. 17 - 43, January 2013

    work page 2013

  4. [4]

    TI Fossen, KY Pettersen and H Nijmeijer, Sensing and Control for Autonomous Vehicles: Applications to Land, Water and Air Vehicles, Springer, 2017

    work page 2017

  5. [5]

    Data association techniques for advanced driver assistance systems using embedded soft-core processors,

    Jehangir Khan, Charles Tatkeu, Pascal Deloof and Smail Niar, “Data association techniques for advanced driver assistance systems using embedded soft-core processors,” 11th International Conference on ITS Telecommunications, pp. 23-25, Aug. 2011

    work page 2011

  6. [6]

    Probabilistic Data Association for Semantic SLAM,

    Sean L. Bowman, Nikolay Atanasov, Kostas Daniilidis and George J. Pappas, “Probabilistic Data Association for Semantic SLAM,” IEEE International Conference on Robotics and Automation (ICRA), Singapore, pp. 1722-1729, May 2017

    work page 2017

  7. [7]

    Machine Learning Methods for Solving Assignment Problems in Multi-Target Tracking,

    PATRICK EMAMI, PANOS M. PARDALOS, LILY ELEFTERIADOU and SANJAY RANKA, “Machine Learning Methods for Solving Assignment Problems in Multi-Target Tracking,” arXiv:1802.06897v1 [cs.CV], Feb. 2018

    work page arXiv 2018

  8. [8]

    Multitarget Tracking,

    Ba-Ngu Vo, Mahendra Mallick and Yaakov Bar-Shalom, “Multitarget Tracking,” Wiley Encyclopedia of Electrical and Electronics Engineering, Wiley, Sept. 2015

    work page 2015

  9. [9]

    An LP-based algorithm for the data association problem in multitarget tracking,

    P. Storms and F. Spieksma, “An LP-based algorithm for the data association problem in multitarget tracking,” Proceedings of the Third International Conference on Information Fusion, pp. 10-13, July 2000

    work page 2000

  10. [10]

    Multitarget Tracking in Clutter: Fast Algorithms for Data Association,

    B. ZHOU and N. K. BOSE, “Multitarget Tracking in Clutter: Fast Algorithms for Data Association,” IEEE Transactions on Aerospace and Electronic System, Vol 29, No 2, pp. 352-363, 1993

    work page 1993

  11. [11]

    Robust approaches for the data association problem,

    Hassene Aissi, Daniel Vanderpooten, Jean Michel Vanpeperstraete. “Robust approaches for the data association problem,” 7th International Conference on Information Fusion, pp. 25-28, July 2005

    work page 2005

  12. [12]

    Linear multitarget integrated probabilistic data association for multiple detection target tracking,

    Yuan Huang, Taek Lyul Song and Da Sol Kim, “Linear multitarget integrated probabilistic data association for multiple detection target tracking,” IET Radar, Sonar & Navigation, , Volume 12, Issue 9, pp. 945-953, 2018

    work page 2018

  13. [13]

    Joint Probabilistic Data Association Filter with Unknown Detection Probability and Clutter Rate,

    Shaoming He, Hyo-Sang Shin and Antonios Tsourdos, “Joint Probabilistic Data Association Filter with Unknown Detection Probability and Clutter Rate,” Sensors, Vol. 18, No. 1, pp. 1-14, 2018

    work page 2018

  14. [14]

    Joint Probabilistic Data Association Revisited,

    Seyed Hamid Rezatofighi, Anton Milan, Zhen Zhang, Qinfeng Shi, Anthony Dick and Ian Reid, “Joint Probabilistic Data Association Revisited,” ICCV, pp. 3017-3055, 2015

    work page 2015

  15. [15]

    Iterative joint integrated probabilistic data association filter for multiple-detection multiple- target tracking,

    Yifan Xie, Yuan Huang and Taek LyulSong, “Iterative joint integrated probabilistic data association filter for multiple-detection multiple- target tracking,” Digital Signal Processing, Volume 72, pp. 232-243, January 2018

    work page 2018

  16. [16]

    Multiple hypothesis tracking for multiple target tracking,

    S. S. Blackman, “Multiple hypothesis tracking for multiple target tracking,” IEEE Aerospace and Electronic Systems Magazine, Vol. 19, No. 1, pp. 5–18, Jan. 2004

    work page 2004

  17. [17]

    Multipath Maximum Likelihood Probabilistic Multihypothesis Tracker for Low Observable Targets,

    X. Tang, Q. Wu, R. Tharmarasa and T. Kirubarajan, “Multipath Maximum Likelihood Probabilistic Multihypothesis Tracker for Low Observable Targets,” IEEE Transactions on Aerospace and Electronic Systems, Vol. 54, No. 1, pp. 502-510, 2018

    work page 2018

  18. [18]

    Smoothing Multi- Scan Target Tracking in Clutter,

    Darko Mušicki, Taek Lyul Song and Tae Han Kim, “Smoothing Multi- Scan Target Tracking in Clutter,” IEEE Transactions on Signal Processing, Vol. 61, No: 19, pp. 4740-4752, 2013

    work page 2013

  19. [19]

    Multidimensional assignment formulation of data association problems arising from multitarget and multisensor tracking,

    A.B. Poore, “Multidimensional assignment formulation of data association problems arising from multitarget and multisensor tracking,” Computational Optimization and Applications, Vol. 3, pp. 27-57, 1994

    work page 1994

  20. [20]

    On the Quality of Optimal Assignment for Data Association,

    Jean Dezert and Kaouthar Benameur, “On the Quality of Optimal Assignment for Data Association,” International Conference on Belief Functions, pp. 374-382, 2014. [21] John J. Hopfield and D. W. Tank, “Neural Computation of Decisions in Optimization Problems,” Biological Cybernetics, Vol. 52, No. 3, pp. 141-152, 1985

    work page 2014

  21. [21]

    Digital Hardware Realization of a Recurrent Neural Network for Solving the Assignment Problem,

    Hung D.L. and Wang J., “Digital Hardware Realization of a Recurrent Neural Network for Solving the Assignment Problem,” Neurocomputing, Vol. 51, pp. 447-461, April 2003

    work page 2003

  22. [22]

    Deep Neural Networks for Linear Sum Assignment Problems,

    Mengyuan Lee et al , “Deep Neural Networks for Linear Sum Assignment Problems,” IEEE Wireless Communications Letters, Vol 7, No 6, pp. 962-965, 2018

    work page 2018

  23. [23]

    Sparse Clustered Neural Networks for the Assignment Problem,

    Saïd Medjkouh et al , “Sparse Clustered Neural Networks for the Assignment Problem,” The Ninth International Conference on Advanced Cognitive Technologies and Applications, pp. 69-75, 2017

    work page 2017

  24. [24]

    Data-Driven Approximations to NP-Hard Problems,

    Anton Milan et al , “Data-Driven Approximations to NP-Hard Problems,” Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, pp. 1453-1459, 2017

    work page 2017

  25. [25]

    Online Multi-Target Tracking Using Recurrent Neural Networks,

    Anton Milan et al , “Online Multi-Target Tracking Using Recurrent Neural Networks,” Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, pp. 4225-4232, 2017

    work page 2017

  26. [26]

    Khizer Amin, Multiple Object Tracking using RNNs, Mastet thesis, Universitat Politècnica de Catalunya, 2017

    work page 2017

  27. [27]

    Deep tracking in the wild: End-to-end tracking using recurrent neural networks,

    Julie Dequaire et al, “Deep tracking in the wild: End-to-end tracking using recurrent neural networks,” The International Journal of Robotics Research, pp. 1-21, 2017

    work page 2017

  28. [28]

    Deep Recurrent Neural Network for Multi-target Filtering,

    Mehryar Emambakhsh et al , “Deep Recurrent Neural Network for Multi-target Filtering,” 25th International Conference on MultiMedia Modeling, pp. 519-531, December 2018

    work page 2018

  29. [29]

    Deep Tracking: Seeing Beyond Seeing Using Recurrent Neural Networks,

    Peter Ondruska and Ingmar Posner. “Deep Tracking: Seeing Beyond Seeing Using Recurrent Neural Networks,” Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, pp. 3361-3367, 2016

    work page 2016

  30. [30]

    Deep neural networks for data association in particle tracking,

    Yao Yao et al, “Deep neural networks for data association in particle tracking,” IEEE 15th International Symposium on Biomedical Imaging, April 2018

    work page 2018

  31. [31]

    Learning to Multi-Targets Tracking in Dense Clutter Environment with JPDA-Recurrent Neural Networks,

    Hui Zhang, Huajun Liu, Cailing Wang, “Learning to Multi-Targets Tracking in Dense Clutter Environment with JPDA-Recurrent Neural Networks,” 3rd International Conference on Control Engineering and Artificial Intelligence, 2019

    work page 2019

  32. [32]

    Multi-Agent Reinforcement Learning for Multi-Object Tracking,

    Pol Rosello and Mykel J. Kochenderfer, “Multi-Agent Reinforcement Learning for Multi-Object Tracking,” AAMAS 2018, July 2018

    work page 2018

  33. [33]

    Collaborative Deep Reinforcement Learning for Multi-Object Tracking,

    Liangliang Ren, Jiwen Lu, Zifeng Wang, Qi Tian and Jie Zhou, “Collaborative Deep Reinforcement Learning for Multi-Object Tracking,” ECCV, pp. 1-17, 2018

    work page 2018

  34. [34]

    Multitarget Tracking via Mixed Integer Optimization,

    Dimitris Bertsimas et al , “Multitarget Tracking via Mixed Integer Optimization,” IEEE Transactions on Automatic Control, Vol. 63, No. 11, pp. 3627-3642, Nov. 2018

    work page 2018

  35. [35]

    A Linear Programming Approach for Multiple Object Tracking,

    Hao Jiang, Sidney Fels and James J. Little, “A Linear Programming Approach for Multiple Object Tracking,” CVPR, June 2007

    work page 2007

  36. [36]

    Tracking Multiple Maneuvering Targets Using Integer Programming and Spline Interpolation,

    Date K. and Nagi R., “Tracking Multiple Maneuvering Targets Using Integer Programming and Spline Interpolation,” 21st International Conference on Information Fusion, pp. 1293-1300, 2018

    work page 2018

  37. [37]

    Learning to forget: continual prediction with LSTM,

    F.A. Gers, J. Schmidhuber and F. Cummins, “Learning to forget: continual prediction with LSTM,” 9th International Conference on Artificial Neural Networks: ICANN '99, pp. 850-855, 1999

    work page 1999

  38. [38]

    Variants of RMSProp and Adagrad with Logarithmic Regret Bounds

    Mahesh Chandra Mukkamala and Matthias Hein, “Variants of RMSProp and Adagrad with Logarithmic Regret Bounds,” arXiv:1706.05507v2 [cs.LG], 28 Nov 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  39. [39]

    A Metric for Performance Evaluation of Multi-Target Tracking Algorithms,

    Branko Ristic, Ba-Ngu Vo, Daniel Clark and Ba-Tuong Vo, “A Metric for Performance Evaluation of Multi-Target Tracking Algorithms,” IEEE Transactions on Signal Processing, Vol. 59, No. 7, pp. 3452-3457, 2011

    work page 2011