Enhancing Multi-Object Tracking with Spatiotemporal Graph Convolution and Trajectory Recovery

Multi-object tracking (MOT) has found widespread application in various fields such as intelligent surveillance, autonomous driving, sports analysis, and drone applications. When video frames are lost or objects are occluded for a long time, the process noise in multi-object tracking based on Kalman filtering increases over time, leading to inaccurate estimations. Firstly, we proposed a spatiotemporal graph convolution interaction module that aggregates interaction information to extract the relationship features of trajectory temporal and spatial information, capturing the temporal and spatial dynamic changes and estimating future movements. Secondly, to alleviate the problem of missed detections, we designed a trajectory recovery module which developed a velocity continuity-based interpolation algorithm. By introducing smoothness constraints, this algorithm reasonably inferred the missing data in both time and space to ensure the continuity and integrity of trajectories. Finally, to reduce trajectory fragmentation, we adopted a multi-matching strategy. Through an adaptive association cost adjustment mechanism, it ensured the effective association of each detection box. The experiments showed that on publicly available benchmark datasets like the MOT Challenge, this approach performed better than existing multi-object tracking networks.