Research On Multi-object Tracking Algorithm Based On Deep Graph Matching

Multi-object tracking task is not only one of the basic research problems of computer vision,but also the basis for high-level semantic analysis tasks such as action recognition and behavior analysis.It has been extensively used in intelligent video monitoring,unmanned driving,military defense and other fields.Growing domestic and foreign experts and scholars pay more attention to it.The purpose of multi-object tracking is to determine the positions of all objects in each frame of the video sequence and keep their identity information unchanged.In recent years,many research works usually combine similarity calculation with heuristic matching solver to solve this problem.This heuristic method makes it difficult to implement an end-to-end training framework,and the similarity calculation methods used in these methods are relatively simple,which leads that the correlation between objects is not fully exploited,thus affecting the accuracy of multi-object tracking.Therefore,this paper proposes an online end-to-end multi-object tracking framework based on graph convolutional neural network.On this basis,a MOT algorithm based on the attention mechanism is proposed.Two algorithms are robust to external conditions and have strong generalization ability.The research contents and results of this paper are as follows:(1)A multi-object tracking algorithm based on graph convolutional neural network is proposed.Existing algorithms usually transform the data association problem in multi-object tracking into a bipartite graph matching problem,and only use the intersection ratio or feature vector similarity between objects to calculate the correlation coefficient,so it is hard to capture the objects dependency relationships between adjacent frames.To tackle this problem,this paper proposes an online MOT algorithm based on graph convolution,which constructs the appearance and position features of the object into a complete graph,and uses the update mechanism of graph convolution network to update the nodes and the edges in the graph to fully capture the global relationship between two adjacent video frames to improve the accuracy of multi-object tracking.(2)A multi-object tracking algorithm based on attention mechanism is raised.The above-mentioned multi-object tracking method based on graph convolutional neural networks heavily relies on manually set object dependency relationships(ie,edge connections in the graph),which are sensitive to the setting of many hyperparameters,and the algorithm robustness needs to be improved.To tackle the problem,an attention framework is introduced in this paper,a self-attention mechanism is used to learn the dependency relationship between inner-frame objects,and a cross-attention mechanism is used to learn the association relationship of inter-frame objects.The algorithm can automatically learn high-quality object dependency relationships,reduce the dependence on manually setting hyperparameters,and upgrade the accuracy and robustness of multi-object tracking.In this paper,two multi-object tracking algorithms are validated on MOT17,MOT16,and MOT15 public datasets separately,and it is verified that the method proposed in this paper can effectively enhance the performance of tractor.