A lightweight scheme of deep appearance extraction for robust online multi-object tracking

Appearance-based Multi-Object Tracking (MOT) methods rely on the appearance cues of objects. However, existing deep appearance extraction schemes struggle to balance speed, performance, and memory footprint. In this article, a lightweight Re-identification network named Fast OSNet is proposed by simplifying the OSNet structure, adding attention modules, and introducing a global and partial-level feature fusion mechanism. To reduce the impact of occlusion noise on trajectory appearance states, the Hierarchical Adaptive Exponential Moving Average (HAEMA) is proposed, which employs adaptive update weights with a two-stage linear transformation. Together, Fast OSNet and HAEMA make up the proposed lightweight scheme. To validate the proposed scheme, it is combined with the full detection-association algorithm BYTE and proposed Fast Deep BYTE Track (FDBTrack). On the MOT17 test set, it achieves 63.2 High-Order Tracking Accuracy (HOTA) and 77.7 Identification F1-score (IDF1). On the more challenging MOT20 test set, it achieves 62.0 HOTA and 75.9 IDF1. It can serve as an auxiliary mean to improve the tracking performance of online MOT methods. The codes are available at https://github.com/LiYi199983/FDBTrack.