Research On Real-time Recognition Technology Of Traffic Signs During Driving

Natural road scene has a large number of traffic sign information,can guide drivers to normal driving.The real-time identification technology of traffic signs plays a great role in advanced auxiliary driving system and autopilot,which is helpful for drivers to understand the information of road congestion in front and guide drivers to carry out standardized driving operation.But in fact,there are too many complex factors in the environment of natural road scene: adverse weather,strong light irradiation,distortion and occlusion will bring great difficulties to the identification of traffic signs.Based on the current YOLO series method of target detection,the following research: First propose the YOLOv4++ algorithm,based on the YOLOv4 algorithm,under the output of the original YOLOv4 model,make full use of shallow features for detection of small targets,output one feature layer in shallow,finally output four feature maps,use four feature maps for prediction,so as to solve the problem of small target leakage detection.Secondly,the K-Means++ algorithm,based on the original K-Means algorithm,generates the prior box on the TT100 K dataset.Based on four output feature maps,12 different prior boxes were obtained on four different scale scales on the grid of the actual natural scene,which makes the prediction box position positioning more accurate and shortens the target detection time to improve the real-time of the algorithm.Finally,the detection scope of traffic signs is expanded,so that the traffic sign identification system can not only identify the signs itself,but also identify the characters inside some signs.To verify the effectiveness of the algorithm improvement,it was compared with the original YOLOv4,YOLOv3,Faster-RCNN.The results showed that YOLOv4++algorithm is 0.96,0.81,0.98 under three natural conditions under three natural conditions;the FPS,average value for 20km/h is 28.3,27.9,26.7 respectively.The improved algorithm has higher detection accuracy,and real-time also improves compared with other algorithms,and can meet real-time recognition to prove the effectiveness of the improved algorithm in this paper.