Research On Detection Method Of Urban Road Traffic Signal Based On Deep Learning

From assisted driving to highly automated driving to unmanned driving,reliable traffic environment perception is the most important and basic in any link.As an important traffic facility that guides the orderly traffic of vehicles at urban road intersections,traffic lights have an important role in improving urban road safety and improving traffic efficiency.However,traffic lights are different from other road elements,and their color information can only be captured by visual sensors.Therefore,the visual detection technology of traffic lights is an indispensable part of traffic environment perception.The major breakthrough in computer vision of deep learning theory and methods has provided a new solution for traffic signal detection.This paper uses the Retinanet model and YOLO V3 model in the deep learning target detection algorithm to detect urban road traffic lights and accurately identify the location and type of traffic lights in urban road images,which can not only lay the theoretical foundation for the arrival of the era of driverless driving.It can also assist the driving of visually impaired people.The main research of this article is as follows:(1)In the traffic signal color detection,this paper based on the Retinanet model from the feature extraction network,network structure and prediction box Anchor preset value three aspects of model improvement research,Retinanet improved model in accuracy rate,recall rate and F1 value and other indicators The performance is superior,indicating that the improvement of the Retinanet model is effective through the three aspects of the feature extraction network,the network structure and the preset value of the prediction box Anchor.(2)In terms of shape detection of traffic lights,this paper is based on YOLO V3 model to optimize its research from four aspects: GIOU,CBAM,Soft NMS and prediction box Anchor preset value.YOLO V3 improves the model in accuracy,recall and The F1 value and other indicators are superior in performance,indicating that the introduction of GIOU,CBAM,Soft NMS and improved prediction box Anchor preset values can help improve the detection effect of the YOLO V3 model.(3)In terms of color and shape fusion detection of traffic lights,this paper proposes two fusion methods.The first is a fusion method based on a feature extraction network.This method is to fuse the shape and color feature map obtained by the feature extraction network.Then the detection of traffic lights.The second is to merge the detected traffic signal shape and color results,and then use the small convolutional neural network to classify the traffic lights.The fusion method based on feature extraction network is superior in indicators such as accuracy,recall and F1 value,which shows that the shape and color feature map obtained by the feature extraction network is fused and then the detection of traffic lights is used to improve the detection effect of traffic lights.effective.