Online Multiple Object Tracking Algorithm Based on Heat Map Propagation

Wei Ding, Haokai Hong, Dejun Xu, Min Jiang

Abstract: Online multiple object tracking is an important computer vision task with a wide range of application scenarios which integrates target detection, re-ID (Re-identification) and feature association matching. Most of the existing methods have two problems: ignoring the spatio-temporal information, resulting in poor tracking effect, and using two-stage detector makes the tracking speed slow. In this paper, we design an online multiple object tracking algorithm by generating a heat map based on the spatio-temporal information of the tracker, and propagate it to a one-stage detector to improve tracking speed and stabilize detection quality. Without introducing additional regressors, the generation of the propagation heat map are both simply and efficiently. Consequently, the proposed algorithm can achieve a balance between the speed and accuracy, and provides a paradigm for the utilization of the spatio-temporal characteristic information of the one-stage detector. Experiments are carried out on MOT-17 and 2DMOT-15 which verifies that 43.27% and 63.7% improvement in tracking speed is obtained with a small accuracy compromise.