Utilizing Category-Specific Re-Identification Models in Drone-Based Multi-Class Multi-Object Tracking

Tracking multiple objects is of high importance for many applications in the field of automated driving or surveillance. Despite the success of re-identification (ReID) models for extracting distinctive appearance features in multi-person tracking, the use of category-specific ReID models in a multi-class tracker has not yet been explored. In this work, ReID datasets for several classes are constructed from the VisDrone dataset, and specialized ReID models are trained on the combination of categories with similar appearance. Next to these ReID models, a sophisticated tracking framework with established components from the literature is built, which greatly surpasses the state of the art on VisDrone-MOT test-dev (+5.0 MOTA, +5.2 IDF1). To meet the runtime requirements of practical applications, a more efficient variant of the tracker is suggested, which runs at 20 frames per second without a significant loss in accuracy.