How have the continuous optimization strategies in Multi-Object Tracking (MOT) evolved in response to the development of object detection technologies, and what are the key findings and applications discussed in the analysis of the last three years' research in this context?
Target detection technology, as the foundation of multi-object tracking, plays a crucial role in improvinging the tracking outcomes. In recent years, benefiting from the continuous advancements in target detection technology, multi-object tracking has experienced rapid development. We will analyze the optimization strategies and key findings in MOT separately from the perspectives of Tracking-by-Detection (TBD) and Joint Detection and Tracking (JDT) methodologies.
Firstly, for Tracking-by-Detection (TBD) approaches, the tracking methods heavily rely on the performance of detectors. A more powerful detector implies stronger tracking capabilities. TBD-based tracking methods combine different detectors, focusing on improving data association. Concurrently, leveraging the trend of deep learning, these methods continually optimize tracking strategies, further enhancing the performance of multi-object tracking. Over the past three years, studies such as MAT, ByteTrack, StrongSORT, C-BIou, BoT-SORT, OC-SORT, DEEP OC-SORT have been discussed. These investigations effectively utilize high-performance detectors to enhance the accuracy of target detection in tracking, optimizing feature extraction, and data association strategies to improve the efficiency and accuracy of trackers.
On the other hand, for Joint Detection and Tracking (JDT) approaches, this methodology integrates detection and tracking frameworks into an end-to-end model, collectively optimizing objectives to enhance the overall system performance. Over the past three years, studies including QDTrack, RelationTrack, CSTrackV2, AttTrack, FairMOT, CorrTracker, SGT, FineTrack have been examined. Embedding "target detection" and "target tracking" tasks into a shared model accelerates MOT and improves target tracking performance. Additionally, by balancing detection tasks and Re-ID tasks, optimizing tracking association strategies, and jointly training detection and tracking tasks, these approaches enhance multi-object tracking effectiveness.
In summary, in recent years, with the rapid development of target detection technology, multi-object tracking has seen continuous growth. Numerous research findings have emerged, with wide applications in fields such as Military, Motion Analysis, Medical Diagnosis, Autonomous Driving, Video Surveillance. The research in this domain holds significant importance and promising applications.
