<p>Existing multi-object tracking (MOT) methods integrate object motion and appearance features to improve tracking performance. However, these methods often exhibit sub-optimal tracking performance when facing objects with diverse poses, similar appearances, or nonlinear motions. Especially in crowded scenarios, the frequent occlusion between objects leads to an increase in the discontinuity of the trajectories, thus lagging the tracking performance. Motivated by this, we propose a heterogeneous feature driven cascade association network (HFCANet). Specifically, we design a triple heterogeneous feature extraction (THFE) module to capture more discriminative features of objects, which can realize accurate object representation even in the presence of drastic changes in scale, pose, and background. In addition, to retrieve the fragmented trajectory of the occluded object for more complete tracking, we introduce an easy-to-hard cascade association policy that analyzes heterogeneous cues to determine whether the detected object can be re-corrected into the trajectory. HFCANet comprehensively considers the heterogeneous cues of objects, encompassing the nonlinear motion and the richness of limb movements. This characteristic makes it especially well-suited for fulfilling tracking requirements in scenes with dense pedestrians and frequent occlusions. Extensive experiments and ablation studies conducted on MOT17, MOT20, and DanceTrack demonstrate the effectiveness of our proposed HFCANet, where ours has achieved better performance compared to existing state-of-the-art (SOTA) methods.</p>