Scalable real-time multi-target tracking and its implementation on DSP

Abstract

While various online multi-target tracking methods have been proposed recently, most of their runtime speed struggle at 1-10 frames per second for moderate crowded scenes. In this thesis, we present a novel real-time multi-target tracking system based on the tracking-by-detection framework. Our system is designed for tracking a variable number of interacting targets from a single, static, above shoulder camera, which is a general setting for video surveillance. One challenge in our approach is that when background subtraction is used for detecting moving targets, merged measurements occur frequently because of target interactions. To cope with the problem, we propose to use correlation filter based object detector to robustly separate the targets in merged measurements. Then, online object tracking assisted data association is used to solve the track-measurement assignment. To reduce computation load, our object tracking algorithm is assisted by correlations filter based trackers which share the same features used by our object detector. In addition, to recover partially occluded targets, we allow unconfident detections to be assigned to tracks whilst care is taken to avoid introducing additional false positives. We also analyze the online approximation to multi-channel correlation filters. Our experiments show that exact solution is more resistant to noisy channels than approximate solution. Evaluation on generally accepted datasets reveals that the proposed system is comparable to state-of-the-art methods in terms of performance while running several magnitudes faster.

Additionally, we show that the proposed system can be readily implemented on the Texas Instruments TMS320C6678 DSP (C6678) without significant degradation in speed or performance. Details on efficient implementation of the system is also discussed. Especially, for computing Histogram of Oriented Gradients (HOG) feature, our optimized implementation runs at 60fps on VGA images on a single core of C6678, which is 10 times faster than a directly ported implementation.