A 3D Object Tracking Method Based on Graph Neural Network and Spatio-Temporal Dependency Knowledge
Yixin Chen, Qingnan Li, Xi Hu · 2023
In road traffic scenes, the personalized motion of the target object, the interweaving motion of the target object and its neighbors, and the interaction and collaboration between the target object and the traffic environment are very complicated. It is difficult to accurately track a target that is heavily occluded or continuously occluded through affinity calculation or associated tracking. In response to this problem, the applicant first proposed a graph neural network based method with multi-feature learning to explore the coordinated change law of the target object's position, direction, and speed, and mine the spatio-temporal dependence knowledge in the coordinated change of the target object's motion, and improve the graph node aggregation feature discrimination. And then the applicant proposes a graph neural network based method with multi-source feature interweaving, exploring the interweaving motion law of the target object and neighboring moving objects, and iteratively updating the graph node features through multi-source feature aggregation to mine the relationship between neighboring data samples. Finally, a graph neural network based method with environmental interaction and collaboration is proposed to explore the interaction and collaboration rules between the target object and the traffic environment. With the design of the target-environment multi-level graph node feature update strategy, the fine-grained multi-source traffic environment facilities are aggregated to mine the knowledge of the temporal and spatial dependence of the interaction between the target object and the traffic environment. Therefore, the proposed method can accurately calibrate and stably track the target object under severe occlusion or continuous occlusion.