Detection and Association Based Multi-target Tracking in Surveillance Video

The Multiple Target Tracking (MTT) problem is one of the fundamental challenges in computer vision. In this paper, we propose a feasible detection and association based MTT system which uses a modified Deformable Part-Based Model (DPM) to generate detection results and then links detections into track lets to further form long trajectories. We first describe our modified DPM algorithm which could automatically discovery optimal object part configurations to improve detection performance. Next to tackle the MTT problem, e.g., Associating detections under imperfect detector identifications, severe occlusions and interferences between objects, etc conditions, we introduce an EM-like inference algorithm that alternatively optimizes the Trajectory Models (TM) for all the targets and the Maximum A Posterior (MAP) solution of the Markov Random Field(MRF) model. At the E-step, we update the TM based on the inference result of the current MRF model, and at the M-step, we use the up-to-date TM to re-compute the probabilities in the MRF model to re-fine the MAP solution. As shown by our experimental results, the presented detection and association based MTT system leads to satisfactory performance.