Usage of Vehicle Re-Identification Models for Improved Persistent Multiple Object Tracking in Wide Area Motion Imagery

Persistent multiple object tracking in Wide Area Motion Imagery (WAMI) is fundamental for a wide range of applications, e.g. surveillance of borders. Though impressive tracking results have been achieved by combining appearance based and motion based object detection as input for modern tracking-by-detection methods, the number of identity-switches (ID-switches) in case of many slow or stopped vehicles is still high. Instead of extracting features from the appearance based object detection model for data association between object detections and existing tracks, we propose to use visual appearance descriptors learned from re-identification tasks. For this purpose, we employed the recent re-identification model OSNet and created an aerial vehicle re-identification dataset based on publicly available aerial object tracking datasets. By applying the optimized data association scheme, we outperform state-of-the-art trackers, as the number of ID-switches is considerably reduced.