AFSTrack: An Approach uses Adaptive Filtering and Spatial Analysis Techniques to Enhance Multi-Object Tracking in Crowded Scenes

The challenges of tracking multiple objects in crowd environments have greatly influenced the practicality and effectiveness of existing tracking algorithms. While many previous algorithms rely on basic Kalman filtering or Noise Scale Adaptive Kalman algorithm, they overlook the influence of system noise. In our proposed approach, we refine the Sage-Husa Kalman filter to mitigate noise interference in crowded scenes, and introduce a straightforward yet efficient data association method for alleviating the problem of target overlap. Similar to ByteTrack, we categorize detections into high confidence and low confidence groups. During the matching process for high confidence detections, we incorporate a relative height cost matrix. For low confidence matching, we employ the EDIOU to expand the target’s spatial coverage. Furthermore, to mitigate target ID Switches caused by occlusion, we associate low confidence detections with inactive trajectories. Finally the ablation experiments demonstrate that our approach remains competitive among top-performing methods, even without utilizing the ReID module.