HRTrack: Enhancing Multi-Object Tracking with Adaptive Multi-Scale Fusion and Occlusion-Aware Attention

Existing joint detection and embedding (JDE) methods achieve promising real-time performance in multi-object tracking, but suffer from limited target association accuracy. This paper proposes a novel JDE paradigm - HRTrack. By integrating a lightweight adaptive feature fusion module and incorporating a decoupled channel attention module as a dedicated ReID branch, HRTrack optimizes detection and target feature extraction jointly. For trajectory association, the LSE-CBIoU tracker, a hierarchical method combining "Lenient Entry, Stringent Exit"strategy with Cascaded Buffered Intersection over Union, achieves optimal matching. The proposed paradiam supports online real-time tracking, attaining 75.4% MOTA and 60.7% HOTA on the MOT17 benchmark.