Research On Vehicle Detection Method Based On Graph Pyramid Transformer

Vehicle object detection technology plays an important role in intelligent transportation system,which lays a solid foundation for the realization of traffic flow information statistics,intelligent parking lot and autonomous driving.In the actual road traffic scene,due to the constant changes of vehicle positions,vehicles often block each other,which makes it impossible to fully extract vehicle feature information in the detection process,resulting in reduced detection accuracy and missing detection.At present,the mainstream vehicle detection algorithms are usually based on Convolutional Neural Networks(CNNs)to extract features,which have high detection accuracy and robustness in complex environments,but the local characteristics of convolutional kernel limit the ability of the network to acquire the features of the global receptive field context.In view of the above problems,the current vehicle detection model is improved and innovated in this study,and a vehicle detection algorithm based on graph pyramid Transformer is proposed.The algorithm integrates the local feature of Graph Convolutional Network(GCN)and the global feature of Transformer.The research results are as follows:(1)This algorithm constructs a feature mixing module.By taking advantage of the graph convolutional network to extract local feature information and Transformer to extract global feature information,a hybrid feature block is constructed that combines the local feature learning capability of GCN network with the global feature acquisition capability of attention mechanism.The hybrid module is composed by serial stacking convolution blocks and attention modules.The feature capture capability of global context based on attention module effectively solves the problem of local receptive field of graph convolutional networks,and further improves the feature extraction capability of the networks.(2)Construct the graph structure representation of the image.The algorithm represents the image in a flexible way.Locality Sensitive Hashing(LSH)algorithm is introduced to establish sparse adjacent node sets,build local and global sparse graphs,and effectively calculate the correlation between nodes in each node set without searching all nodes in the graph.The computational complexity of attention mechanism O(n~2)is effectively alleviated.(3)A bidirectional pyramid feature fusion module is designed by improving the feature pyramid network.Due to the problem of vehicle scale change,small vehicles may miss detection.In this paper,a bidirectional multi-scale feature fusion structure is constructed to fully integrate the shallow position information and deep semantic information extracted from the backbone network,so as to enhance the information representation ability of feature maps and improve the recognition ability of the model to small objects.