Data-driven Physical Adjustment for Nonlinear Motion Tracking

Abstract

Multiple object tracking in real-world pedestrian scenes is often degraded by false positive detections and unstable prediction-detection alignment arising from object proximity, partial occlusion, and overlapping motion. These factors can disrupt data association and lead to inaccurate trajectory estimation. To address these issues, we propose a two-stage, data-driven approach based on physical correction to improve tracking stability. At the detection stage, we effectively remove false positive bounding boxes by actively searching for candidates relative to target objects by using a rule-based filtering process. At the tracking stage, we improve the intersection over union (IoU) scores of active tracklets against detection boxes based on the conditional adjustment of offset values in the residual covariance matrix while updating states via Kalman filter predictions that account for the nearest neighboring tracked objects. According to our experiments on the Multiple Object Tracking (MOT)17 dataset, the data-driven physical adjustment approach demonstrates a significant reduction in false positives from 3381 to 3305 while maintaining the recall score. When both components are applied together, the tracker achieves the best overall performance, yielding the MOT accuracy, precision, and identification F1-score values of 0.5657, 0.7992, and 0.6446, respectively. Furthermore, the R-offset improves the average IoU of prediction-detection pairs from 0.8691 to 0.8748 across all frames and from 0.8906 to 0.9024 for R-offset-activated pairs. These results suggest that the proposed approach enhances tracking stability by suppressing false positive detections and selectively mitigating prediction errors under proximity and overlap conditions.