Research On Traffic Sign Detection And Recognition Method In Complex Scenes

Traffic signs contains important road indication information.Due to the complexity of vehicle driving roads and a wide variety of signs,the traffic sign detection and recognition has become a hot issue in the research of intelligent transportation and driverless driving.Traffic signs are exposed to the outdoor environment for a long time and are prone to damage and obscuration,thus making the traffic sign detection and recognition system less effective and not better able to satisfy the driver’s required driving assistance information.Based on the YOLOv4 algorithm and the capsule network,this thesis realizes the detection and recognition of traffic signs,and has done the following work:(1)Aiming at the problem of the small targets for traffic sign detection is not effective,a small target traffic sign detection algorithm for YOLOv4 with fused the multiple features is proposed.Based on the network of YOLOv4,the horizontal and vertical double-span layer connection is carried out.Through the dense connection method,a new branch of the shallow network information is vertically added to the feature fusion structure to improve the model’s feature representation ability for small targets.Through the residual connection method,the information of convolutional layer in the front and back parts of the network is fully overlapped to further improve the detection effect of the model on different sizes overall.The algorithm is validated on the GTSDB(German Traffic Sign Detection Benchmark)and CCTSDB(CSUST Chinese Traffic Sign Detection Benchmark)respectively,the experimental results show that the improved YOLOv4 model has better detection effect on small-target traffic signs.(2)Aiming at the problem that the traffic signs are blocked in the complex scenes,resulting in low recognition accuracy,a traffic sign recognition algorithm based on compound capsule network is proposed.The multi-scale idea of residual network is introduced into the convolution layer of capsule network,and the dual channel pooling is added to primary capsule layer.At the same time,the calculation method of dynamic routing algorithm is optimized to improve the effect of feature extraction and the output probability of prediction values,which enables the accurate recognition of blocked traffic signs.The algorithm is validated on the GTSRB(German Traffic Sign Recognition Benchmark),the experimental results show that the improved compound capsule network has a recognition accuracy of 99.21%,and the average accuracy improved by 6.54%,thus the effectiveness of the improved algorithm is verified.(3)A prototype system for traffic sign detection and recognition is developed.The system uses Python language to write code,and the interaction is realized by binding the functions of detection and recognition with the controls in the graphical user interface.The system is easy to operate.