Multi-angle Vehicle Recognition Algorithm And Application Research

In recent years,the supply of China’s transportation system has shown a trend of mismatch with the increasingly growing demand for transportation.Building an intelligent vehicle infrastructure cooperation system with intelligent perception technologies as its core can effectively improve the efficiency of traffic management and operation,which is the core and key of the "new infrastructure" in the new era.However,in real-world complex environments,the effectiveness of vehicle recognition algorithms is significantly reduced,due to changes in factors such as lighting,scenes,shooting angles,and distances.In order to improve the effectiveness and robustness of vehicle recognition algorithms under complex road conditions,this thesis conducts in-depth research on the following aspects:(1)Due to the subtle differences in global appearance features of vehicles at different shooting angles,it is difficult to accurately characterize the vehicle type.Therefore,this article proposes a multi-angle vehicle recognition algorithm based on detail awareness.This algorithm is based on the YOLOX model and constructs a local perception module by introducing the ECA attention mechanism,enhancing the network model’s perceptual ability for the appearance features of different vehicles.At the same time,the Atrous Spatial Pyramid Pooling(ASPP)is introduced to solve the problem of loss of detailed information caused by the increase of receptive field in spatial pyramid pooling(SPP)due to pooling operations.The experimental results show that the accuracy of multi-angle vehicle recognition in actual road’s conditions can reach 96.1% using this algorithm.Compared with the YOLOX model,the recognition accuracy is improved by 6.48%,demonstrating ideal effectiveness and robustness.(2)In order to further improve the expression ability of the YOLOX model for the appearance features of multi-angle vehicles,an improved model based on deformable convolution and multi-order feature interaction is proposed.The model first extracts the main features based on the YOLOX model and refers to the introduction of deformable convolution operations in the two upsampling layers of the feature pyramid network(FPN).This helps to capture features that are more closely related to the target and therefore extract more expressive appearance features..Finally,a recursive gate convolution is introduced in the path aggregation network to achieve multi-order feature interaction,enhancing the network model’s selection of vehicle appearance features.Experimental results show that the proposed improved model achieves further improvement in vehicle recognition accuracy in the same actual road conditions,reaching 96.9%.(3)In order to further improve the efficiency of the vehicle recognition algorithm on massive datasets and facilitate large-scale deployment applications,a lightweight vehicle recognition model based on adaptive feature fusion is proposed.The model replaces the CSPDark Net53 backbone network with the Mobile Net V3 lightweight network,reducing the number of model parameters.At the same time,an adaptive spatial feature fusion module is introduced after the path aggregation network,which enables autonomous learning of the weights of different network layer features to better adapt to the needs of multi-angle vehicle recognition.Experimental results show that the lightweight vehicle recognition algorithm constructed based on the adaptive feature fusion achieves detection speeds of 65 FPS,38FPS,20 FPS,and 34 FPS faster than the classic Fast R-CNN,SSD,YOLOv5,and YOLOX algorithms,respectively,without sacrificing accuracy.