Research On Multi-target Tracking Algorithm Based On Graph Network Optimizatio

Multi-object tracking technology is an essential research direction in the field of computer vision,involving statistical analysis,machine learning,pattern recognition and other disciplines,and is widely used in automatic driving,video surveillance and other areas with significant research value and application prospects.The goal of multi-object tracking is to detect and position multiple objects in a single video sequence,as well as to maintain their identity information in subsequent frames and link their trajectories at last.Multi-object tracking in complex scenes is still challenging,as there exists problems such as similar object epistasis,crowded scenes,and object occlusion.The main bottleneck of these issues is the effective learning of object discriminative features and the validity analysis of data association.This thesis conducts research in two aspects including object feature enhancement and similarity measurement,mainly based on intra-frame relationship modeling and self-attention mechanism to fuse local and global features of objects.In the data association phase,this thesis employs a graph matching module to combine the topological relationships between individual object features and interactions among multiple objects for similarity calculation and optimal assignment to improve the performance of multi-object tracking.The main work of this thesis is as follows:(1)For the object similarity representation problem,this thesis designs a new multi-object tracking approach based on intra-frame modeling and self-attention fusion scheme.When the objects are apparently similar and the motion trajectories are also thundering,it will lead to the situation that both objects in the trajectory and the same object in the candidate frame have high similarity,and it is impossible to distinguish them accurately.For this reason,this thesis build a topological graph for the object and its two nearest neighbors in the frame,and aggregate the features of the surrounding neighbors to enhance its own distinguishability,i.e.,the local features of the object.In addition,this thesis also considers building a bipartite graph between frames to aggregate the feature information of objects in neighboring frames to obtain their global feature representation,and introduces a graph attention mechanism to learn the weight information of local and global features autonomously and fuses them both to obtain the final feature representation of the object.Extensive comparison experiments verify the effectiveness of the intra-frame model and the fusion mechanism.(2)For the problem of object occlusion in complex backgrounds,a multi-object tracking method based on improved learnable graph matching is designed in this thesis.The occlusion problem of known objects will lead to the inability to extract the corresponding effective feature representation,and seriously affect the performance of subsequent object recognition and object correlation analysis.For this reason,we introduces a graph matching module,which also considers the topological relationship between the object and the neighbors during data association,that means we calculate their feature similarity and topological relationship similarity for trajectory objects and candidate objects at the same time.To solve the problem of high computational cost of topological relations,edge filtering strategy and minimum spanning tree algorithm are used to improve,which reduces computational cost while ensuring tracking performance.A large number of comparative experiments verify the effectiveness of the improved learnable graph matching module.(3)Lastly,this thesis has implemented a multi-object tracking system for the models proposed in the previous two approaches.The system has several functional modules,such as user login and registration,model training,model tracking and visualization modules,where the training and tracking processes can be dynamically displayed with visualization functions.