A Simple Multi-Frame Fusion Baseline For Long-Term Multi-Object Tracking
Junmin Ke, Shengting Guo · 2020
Multi-object tracking (MOT) is an important and practical task related to both surveillance systems and moving camera applications. However, little attention has been paid to keep the stability of tracker in long-term multiple object tracking. In this work, we study the essential reasons behind the failure and accordingly propose two techniques to improve the efficiency of MOT: (1) In order to reduce identity switches (IDS) in tracking, we propose a novel approach by accomplishing the detection and re-identification tasks in a single network, which can exploit historical information and generate more robust frame predictions. (2) For the purpose of increasing the accuracy rate of data association, we also design and train an affinity network to measure the object similarity instead of traditional measurements. Overall, experimental evaluations show that our approach achieves excellent performance, especially under the occluded condition and in the case of long-term tracking, as compared to other state-of-the-art methods.