Research On Object Detection Algorithm For Traffic Scenes Based On YOLOv4 Optimization

Image-based traffic object detection is a key technology for intelligent transportation systems,and is of great significance for the intelligent management of traffic.In complex traffic scenes,small-scale and occluded traffic objects are prone to the problems of missed and false detection,and the detection accuracy needs to be improved.The paper improves the YOLOv4 algorithm to enhance the accuracy of traffic object detection in complex traffic scenes.The main research elements are:(1)The AO-YOLOv4(Attention Optimization YOLOv4)algorithm is proposed for the small object detection problem.(a)Using the Canopy algorithm and K-means algorithm to cluster the bounding boxes,the clustered prior frames are more accurate and more representative of multi-scale objects,which effectively improves the accuracy of small object regression;(b)Using Focal-EIo U(Focal and Efficient Io U)to construct the loss function of the bounding box regression,which accurately describes the relationship between the predicted frame(c)Combining the spatial attention mechanism and the channel attention mechanism,the dual-channel attention mechanism module MSCAM(Mixed Spaces And Channels Attention Module)is used to focus the network on the key features and improve the learning of small objects.information and improve the attention to small objects.Experimental results show that the above improvements effectively increase the number of small objects detected and improve the detection accuracy of small objects,with a 1.62% increase in m AP.(2)The improvement method is proposed to address the problem that occluded objects are easily missed and mis-detected.Firstly,the hybrid cavity convolution residual module res HDC(Residual Hybrid Dilated Convolution)is used for feature enhancement,which has a multibranch parallel structure and obtains a multi-scale perceptual field fused feature map by aggregating cavity convolution kernels with different expansion coefficients,effectively enhancing the learning capability of the network for obscured objects;secondly,the Soft-DIo UNMS non-maximum suppression algorithm is used to screen out redundant prediction frames in the post-processing process,which improves the localization accuracy of the occluded objects.The experimental results show that the above optimization strategy effectively improves the problem of missed and false detection of occluded objects,and the m AP increases by 3.03%.(3)A traffic object detection prototype system was designed and implemented,which mainly consists of four parts: user authentication module,data reading module,detection module and result display module,and is capable of real-time and accurate detection of static traffic scene pictures and dynamic traffic videos.