Research On Traffic Signs Detection And Recognition Algorithm Based On Deep Learning

In the field of automatic driving,the accurate detection and recognition of traffic signs is a key technology to ensure the realization of automatic driving in natural scenes,and has very important research significance.The images collected by the on-board camera in the automatic driving system are generally high-resolution panoramic images.The traffic signs occupy a small size in the image.In the general target detection model,the model detection effect is poor.In this paper,the traffic signs in the car camera images are taken as the research object,and the problems of the small size,low detection rate,and poor real-time performance of the traffic signs in the high-resolution images are studied.A series of problems on the mobile terminal embedded devices.The main research content and work are as follows:(1)In view of the small traffic sign missed detection problem in the YOLOv3 network model,the traffic signs detection network model was redesigned and implemented,and the Traffic Signs-YOLOv3(abbreviation: TS-YOLOv3)network model was proposed.The weighted multi-scale feature fusion network improves the detection accuracy of traffic signs and effectively solves the problem of missed detection of smaller traffic signs.(2)The calculation method of the model prior frame clustering algorithm is improved.The K-Means++ clustering algorithm is used to obtain 9 prior frames suitable for the detection of traffic sign data sets.The data enhancement method is used to enrich the traffic sign data sets and enhance the robustness of the convolutional neural network model.(3)Research on the post-processing algorithm of the YOLOv3 network model.When traffic signs overlap in the image,the NMS(Non-Maximum Suppression)algorithm preferentially selects the candidate frame with the highest classification confidence and removes the overlapping Candidate box,only one of the overlapping traffic signs can be detected.The NMS algorithm is changed to the Soft-NMS algorithm.By reducing the classification confidence of overlapping candidate frames,and retaining candidate frames with different targets that overlap to a certain extent,the problem of missed detection caused by the overlap of traffic signs in the image is solved.(4)In order to solve the problem that the TS-YOLOv3 network model obtained by training has a large memory footprint and requires GPU accelerated calculation to achieve real-time detection,the accelerated calculation is achieved by combining the convolution layer and the BN layer,and the model is cut The branch algorithm compresses the network model volume.The weight of the TS-YOLOv3-Prune network model obtained after pruning is reduced by 33 times,the recognition speed is increased by 10 times,and the recognition accuracy can be kept basically unchanged.The experimental results show that the proposed method solves the problem of missing detection caused by the low proportion of traffic signs in the images collected by the vehicle camera,and achieves the model compression.A lightweight traffic signs detection model with equivalent accuracy,simplicity and good real-time performance is obtained.