Multi-Object Tracking Using Video Sequences with Improved Performance and Reduced Computation Cost

Multi-object tracking (MOT) in video sequences is a fundamental problem in surveillance applications, aimed at robustly tracking multiple objects over time in video sequences. Following the tracking by detection framework, in this work, two efficient techniques are introduced to improve the tracking accuracy and reduce the computation complexity. Specifically, a more realistic motion model is used to improve the tracking performance with no extra computation cost, and a novel reduced-order Kalman filter is introduced to reduce the computation cost but with the same tracking accuracy compared to the full-order Kalman filter. Experimental results on MOT datasets show that the proposed method brings competitive results in terms of tracking accuracy and inference speed. In addition, the proposed techniques can easily be migrated into other trackers to improve their tracking performance.