Multi-target tracklet stitching through network flows

Complex track stitching problems have risen in prominence with the development of wide area persistent sensors for urban surveillance. Common approaches such as multiple hypothesis algorithms in tracking and tracklet stitching solve these problems by creating a representation of all possible data associations, then solving for the optimal data association via integer programming or suboptimal techniques. The number of data association hypotheses grows exponentially with number of objects and time, which leads to scaling issues when tracking targets in high-density environments over long periods of time. State-of-the-art multiple hypothesis approaches make tradeoffs to limit complexity, discarding potential data associations to avoid scaling problems. These approaches run the risk of eliminating the best data association hypotheses. This paper describes a novel graphical approach to data representation, called the Track Graph, which provides a compact and efficient structure for the storage of tracklet data and subsequent formulation of the tracklet stitching problem. The Track Graph implicitly represents the set of feasible tracklet stitching hypotheses, with linear scaling in time in both edges (feasible associations) and nodes (tracklets). Using the Track Graph, we pose the track-stitching problem as a min-cost flow problem, and describe polynomial-time algorithms to solve the full tracklet-to-track assignment problem, obtaining the same optimal solution as multiple hypothesis tracking algorithms. We demonstrate the efficacy of our algorithms on high density simulated scenarios.

[1]  Chris Stauffer,et al.  Learning to Track Objects Through Unobserved Regions , 2005, 2005 Seventh IEEE Workshops on Applications of Computer Vision (WACV/MOTION'05) - Volume 1.

[2]  Donald Reid An algorithm for tracking multiple targets , 1978 .

[3]  Nathan Cooprider,et al.  Efficient multiple hypothesis tracking by track segment graph , 2009, 2009 12th International Conference on Information Fusion.

[4]  C. Morefield Application of 0-1 integer programming to multitarget tracking problems , 1977 .

[5]  S. Mori,et al.  Tracking and classifying multiple targets without a priori identification , 1986 .

[6]  David K. Smith Network Flows: Theory, Algorithms, and Applications , 1994 .

[7]  Yiwei Wang,et al.  Moving object tracking in video , 2000, Proceedings 29th Applied Imagery Pattern Recognition Workshop.

[8]  D. Castañón Efficient algorithms for finding the K best paths through a trellis , 1990 .

[9]  Yaakov Bar-Shalom,et al.  Multitarget-Multisensor Tracking , 1995 .

[10]  Chee-Yee Chong,et al.  Generalized Murty's algorithm with application to multiple hypothesis tracking , 2007, 2007 10th International Conference on Information Fusion.

[11]  Aubrey B. Poore,et al.  A Lagrangian Relaxation Algorithm for Multidimensional Assignment Problems Arising from Multitarget Tracking , 1993, SIAM J. Optim..

[12]  Richard Perry,et al.  A NEW PRUNING/MERGING ALGORITHM FOR MHT MULTITARGET TRACKI NG , 2000 .

[13]  Chee-Yee Chong,et al.  Evaluation of Data Association Hypotheses: Non-Poisson I.I.D. Cases , 2004 .

[14]  Jack K. Wolf,et al.  Finding the best set of K paths through a trellis with application to multitarget tracking , 1989 .

[15]  Ramakant Nevatia,et al.  Global data association for multi-object tracking using network flows , 2008, 2008 IEEE Conference on Computer Vision and Pattern Recognition.

[16]  Richard Ivey,et al.  Long-duration fused feature learning-aided tracking , 2010, Defense + Commercial Sensing.