PE-Track: Enhancement of Holistic Motion Prediction for Multiple Object Tracking

Robust and accurate motion prediction is essential for reliable multiple object tracking (MOT). Tracking-by-detection (TBD) methods with a learnable motion predictor exhibit notable performance in handling complex scenarios. However, most of these methods generally neglect the spatial interactions between objects. Furthermore, both detector localization errors and incorrect associations introduce noise into the motion predictor. To address unreliable motion prediction caused by spatial information deficiency and noise propagation, we propose PE-Track, a holistically enhanced tracker. Specifically, we develop a historical–spatial motion predictor (HSMP) to achieve accurate prediction. Each layer of the HSMP employs two independent branches to extract historical motion and spatial interaction features, and these features are mutually guided and enhanced via a fusion block. Moreover, to reduce the matching noise caused by incorrect associations, we introduce a hybrid offset modulated IoU (HOMIoU) to optimize the discriminability of the conventional IoU. It combines the center offset and depth offset to modulate the IoU, sufficiently utilizing the latent information in the prediction results. Additionally, we propose an adaptive position correction update (APCU) strategy to filter the measurement noise caused by detectors. The APCU uses the predicted positions to dynamically adjust the observed positions according to the confidence of these positions. In the experiments, our proposed PE-Track achieves state-of-the-art performance on benchmarks such as Dancetrack and SportsMOT.