RetryTRACK: Recovering Misses in Multi-Camera 3D Pedestrian Tracking

Isabella de Andrade, João Paulo Lima, Veronica Teichrieb

Abstract: Tracking pedestrians commonly relies on detection algorithms. However, these algorithms are not always correct and may miss some pedestrians. Although using multiple cameras is a way to handle this, some failures still need to be fixed. Thus, it is desirable that the tracker attempt to fix the detections. This work introduces an online and untrained module designed to recover missing detections during multiple camera tracking. The module applies linear extrapolation and Gaussian process regression techniques to produce new smoothed coordinates. Furthermore, we propose a filter to remove duplicate detections. We attached the module to a multi-camera baseline tracker and evaluated it on the WILDTRACK and MultiviewX datasets. The multiple object tracking accuracy was improved by 0.95 p.p. on WILDTRACK and 2.34 p.p. on MultiviewX with the addition of the module. Moreover, this strategy successfully recovered 17.98% of missing detections in WILDTRACK and a significant 40.12% in MultiviewX, underscoring its practical application and potential to address the problem.