Research On Vehicle Detection And Recognition Algorithm Based On Deep Learning

With the rapid development of China’s economy,the number of private cars is increasing year by year,which makes the pressure of traffic supervision increasing day by day.Urban traffic problem has become a problem we must solve.In recent years,the rapid development of artificial intelligence has brought more efficient solutions to alleviate the pressure of urban traffic on society.Vehicle detection technology is undoubtedly the key technology in the construction of intelligent transportation,which can save more time and energy for traffic managers.It can not only meet the requirements of real-time detection,but also accurately identify the vehicle information.With the continuous development of deep learning,the research of vehicle detection and recognition using deep learning has become a new development trend.Therefore,this paper uses the object detection algorithm in the field of deep learning for vehicle detection and recognition.The innovation content of this paper is as follows:(1)Due to the limited feature extraction ability of the shallow feature layer in SSD network model,small-scale vehicles are missed.A feature enhancement module is designed to improve the feature extraction ability of the network.In conv4＿3 and FC7 feature layers add feature enhancement module,use 1×1 and 3×3 convolution to extract more features,at the same time,hole convolution is used to expand the receptive field of feature layer,so that the network can learn multi-scale features,and improve the missing detection situation of SSD algorithm in small-scale vehicle detection and recognition.(2)SSD network model is not sensitive to small-scale vehicles because of its weak feature expression ability in shallow feature layer.The feature fusion module is designed to improve the feature expression ability of shallow feature layer.In the feature fusion module,Conv4＿3 and FC7,FC7 and Conv8＿2、Conv8＿2 and Conv9＿2,The fusion method is bilinear interpolation with low time complexity.Enhance the shallow feature layer of small-scale vehicle feature expression ability,improve the detection accuracy of vehicle detection and recognition.(3)In view of the fact that the proportion setting of SSD prior box is not suitable for vehicle detection and recognition task,K-means clustering algorithm is used to cluster the real box annotation information of KITTI data set,and the prior box ratio which is most suitable for vehicle detection and recognition task is obtained,so as to improve the accuracy of target annotation in vehicle detection and recognition.