Research On Vehicle Detection Technology Based On Improved YOLOv5 Network

Vehicle detection as a popular direction in the field of target detection has been fruitful so far,but in actual traffic scenarios,the complex and changing environment and the different scales of vehicle targets can cause the detection process to be unable to fully extract the relevant feature information of the target,resulting in target misses and false detections,which in turn leads to a reduction in the detection accuracy of the algorithm.To address the above problems,this paper takes YOLOv5 as the benchmark model and aims to improve the target detection accuracy,and proposes three improvement strategies,which are mainly as follows:(1)An algorithm YOLOv5＿DA based on an expanded convolutional block attention mechanism is proposed for the problem of noise in images,background interference and no attention preference of the original network.DA,an expanded convolutional block attention mechanism based on CBAM(Convolutional Block Attention Module),is added to the Neck network to enhance the sensory field of the network while reinforcing important features of the target from both channel and spatial dimensions,improving the representational capability of the network model and enabling the fused feature map to contain more valid information,thus improving the detection accuracy of the vehicle.The experimental results demonstrate that YOLOv5＿DA achieves an average accuracy of 79.7%,which is a 1.4% improvement over the original algorithm.(2)An algorithm YOLOv5＿T based on Vi T(Vision Transformer)is proposed for the problem of convolutional neural networks cannot obtain global information and have poor generalization ability.Replaces the input to Vi T with the feature map output from the convolutional neural network and removes the Normalization layer from Encoder,avoiding the risk of corrupting the features extracted by the convolutional operation and the model failing to converge.The problem of poor detail extraction of Vi T structures is improved.The global modelling of convolutional neural networks and the ability to mine long-range dependencies are effectively enhanced to address the need for detection of vehicle targets at different scales.The experimental results demonstrate that YOLOv5＿T achieves an average accuracy of 79.8%,an improvement of 1.5% on the base network.(3)To address the problem that deep convolutional neural networks cannot make full use of shallow information,an algorithm based on adaptive spatial feature fusion,YOLOv5＿ASFF,is proposed.The adaptive spatial feature fusion module is added to the feature pyramid module,and the GIo U is used instead of the Io U loss function for target prediction.The inconsistency between different feature scales is suppressed by filtering conflicting information in the space,and the fusion of feature information at different scales is enhanced,thus improving the vehicle target detection performance.The experimental results demonstrate that the average accuracy of YOLOv5＿ASFF is 80.3%,which is a 2% improvement compared with the original algorithm.Finally,the three improvement strategies are combined to design the YOLOv5＿Ms network,whose detection accuracy is 2.8% higher than that of YOLOv5.In summary,the various improvement strategies proposed in this paper around YOLOv5 are effective in improving the target detection accuracy.