Research On Small Size Traffic Sign Detection And Recognition Algorithm

Traffic sign detection and recognition is an indispensable technology in the intelligent transportation system,which can provide vehicle safety warning,plan traffic paths and improve the transportation efficiency of the road network.Although this detection technology has been studied for many years,with the continuous advancement of intelligent transportation,how to efficiently and quickly detect small traffic signs in complex environments is still faced with many problems.Moreover,due to the limited size of the model,it is difficult to achieve the balance between detection accuracy and detection speed when deployed on small mobile devices.In this regard,the thesis makes use of deep learning technology to carry out in-depth research on traffic sign detection and recognition algorithm.The specific work is as follows:(1)Aiming at the low detection rate of small-sized traffic signs,a traffic sign detection network based on YOLOv5 sampling optimization is constructed.First,use the self-convolution operator Involution to perform self-convolution sampling on the feature map,extract features flexibly and efficiently,and construct a cross-stage attention mechanism(Attention Module Cross Stage Partial,AMCSP)structure to add importance weights to channels,so that The network pays more attention to objects of small size.Then,an improved channel aggregation structure(Path Aggregation Network,PAN)is used to achieve fusion and enhancement of multi-scale semantic information and detailed features.Finally,the DIOU＿NMS function is introduced for post-processing to avoid false suppression of occluded targets and improve target positioning accuracy.The experimental results on the data set of traffic signs in China show that the average accuracy of the network is 95.8% when the threshold of intersection ratio is 0.5.(2)In order to solve the problems of large number of traffic sign detection network parameters and difficult mobile deployment,a lightweight single-stage detection network TSDNet-tiny based on Anchor free is constructed.The network can reduce memory access cost by integrating network branches.Subsampling channels are added to integrate information of different scales to achieve accurate detection of small size targets.Then,ATSS algorithm was introduced to balance positive and negative sample backgrounds,and Focal loss function was used to achieve fast convergence.Mobile phone deployment was carried out through NCNN framework,with 0.94 M network parameters deployed and an average detection speed of 25 FPS.(3)In order to further improve the performance of traffic sign recognition network and distinguish the detailed categories of optimized traffic signs,a recognition method based on the combination of Context enhancement mechanism and self-convolution operator is proposed,which adopts Context Dilated Pooling Pyramid,CDPP structure enhances the ability of the network to obtain multi-scale information,expands the receptive field and improves the robustness of the network when the convolution kernel parameters are unchanged.Secondly,multifold down sampling operation is carried out on the input image by self-convolution operator to flexibly extract features and obtain more abundant image details and semantic information.Experimental results on the data set of traffic signs in China show that the recognition accuracy of this method is 99.12%.