Research On Key Technologies Of Traffic Sign Detection And Recognition Under Complex Urban Scenes

Traffic sign detection and recognition is a gordian technique of the smart car.Because the urban traffic scene is complex and changeable and the model has a bad anti-interference capacity,the traditional model has obvious space for improvement in reliability,real-time performance and stability,which can not meet the requirements of practical application.This paper studies the complex urban scene,the detection and recognition of traffic signs,and the performance indicators of the model,proposing a key technology of traffic sign detection and recognition under complex urban scenes.The main research work is as follows:1.Designing an image preprocessing algorithm based on multi-mode scene fusion.Training a VGG16 neural network to recognize multi-mode scenes(foggy scene,night scene and normal scene)of road images,and finishing the image preprocessing,multi-scene features are integrated to increase the reliability of traffic sign detection.Proposing an improved Mosaic data enhancement algorithm to solve the problem of traffic sign loss caused by traditional Mosaic data enhancement on road images,which provides a basis for traffic sign detection.2.Proposing a multi-scale traffic sign detection algorithm based on improved YOLO＿V4.Using the K-means+ + clustering to generate anchor boxes conform with traffic signs,designing a CSPDark Net33 backbone with a significantly reduced of weight calculations to optimize the structure of YOLO＿V4 neural network,using CIOU Loss,Focal Loss and balanced cross entropy to optimize the YOLO＿V4 Loss function,detecting traffic signs of different sizes through three branches,and increase m AP to 95.45% and FPS to 27.3.Designing a two-step detection and recognition algorithm combined with DCGAN neural network.Using DCGAN neural network to enhance the traffic sign recognition dataset,balance the number of traffic signs and enrich the overall feature information.Designing an improved VGG16 neural network to train three models to recognize warning,prohibition and mandatory respectively.Traffic sign recognition should be further completed in according to the results of traffic sign detection,and the consequence of MAP is 94.02% and FPS is 27.Finally,the three works of multi-mode scene fusion image preprocessing,traffic sign detection and traffic sign recognition are integrated into one flow,on this basis,finishing the video detection and recognition function,and realizing a traffic sign detection and recognition system.Through the actual test of the system,we certify it takes into account the reliability,real-time performance and stability,and further proves the significance of the above research on traffic sign detection and recognition.