Optimisation of occlusion data based on graph neural networks with joint objectives

This paper proposes a data federation multiobjective tracking (MOT) method based on a graph neural network framework, where a graph-based approach merges the optimisation data between multiple objects in the same frame and between objects in the preceding and following frames into the feature learning task. This approach uses the data logic and topological properties of graphs, combined with detectors to encapsulate the neighbourhood information of each node, in order to construct an association graph between multi-objective multi-features with the inference logic of optimised occlusion data, which is then learned by linear filtering to constrain feature score similarity. Finally, this paper conducts comparative experiments on the MOT dataset. By comparing the experimental index data of MOT15/17/20, the outstanding performance of this paper’s method can be effectively demonstrated.