Vehicular/Non-Vehicular Multi-Class Multi-Object Tracking in Drone-Based Aerial Scenes

Despite recent advancements, Multi-Object Tracking (MOT) remains a difficult task, particularly when dealing with drone videos. A variety of challenges, such as data association, varying levels of occlusion, camera motion, and missing detection, affect the effectiveness of MOT tasks. Data association in tracking-by-detection approaches plays a critical role in determining how detections from previous frames are linked to generate coherent tracks. To address this challenge, merging spatial and visual cues may not lead to identical matching pairs for the given detections and tracks. Thus, the performance of the MOT task can be improved by pairing them individually and then determining the best matching pair. It is therefore proposed a Multi-Class MOT (MCMOT) tracker with a cascaded data association strategy in which, during the first stage, matching pairs based on appearance and spatial information are identified separately. The best matching pair is then identified out of these pairs in the second stage and updated as an active track. In addition, the Kalman Filter (KF) estimation and Constant Velocity Assumption (CVA) model are employed to effectively address scenarios involving missed detections and occlusion handling by utilizing motion data. Experiments were carried out on the two state-of-the-art datasets, VisDrone-MOT and UA-DETRAC, to validate the effectiveness of the proposed MCMOT tracker, which demonstrated the improved performance when compared against six different state-of-the-art trackers by providing as input different detectors.