Research And Application Of Traffic Signal Light Detection And Recognition Under Complex Background

Traffic signal light detection and recognition is one of the key technologies in the field of unmanned driving and auxiliary driving.Accurate access to traffic light information in actual road scenes can provide support for drivers or intelligent vehicles to reduce operational errors and correct decisions,and reduce the occurrence of traffic accidents and congestion.Traffic signal detection and recognition is image processing and a research hotspot in the field of intelligent transportation,as a result of the actual traffic intersection in the scene environment is more complex,traffic lights,color changes and easy affected by external factors such as illumination and complex scenes,lead to the existing relevant methods to test the traffic lights and recognition can’t well meet the requirements of real-time and accuracy,Therefore,it is of practical significance to carry out the research on the detection and identification method of signal lamp under complex background.In this paper,image processing and deep learning are combined to detect and identify the image.The main research work is as follows:(1)In order to eliminate the influence of complex background,this paper proposes a method to extract candidate regions of signal lamps.Firstly,the traffic signal light image is preprocessed,including setting up the region of interest(ROI)and histogram equalization image enhancement method to reduce the calculation amount and the influence of illumination change.Secondly,the candidate regions were preliminarily determined by morphological top-hat operation and threshold segmentation based on HSV color space.Finally,combined with the statistical results of the geometric feature data of the traffic signal lamp,the marked connected region is filtered to accurately detect the candidate region of the traffic signal lamp.(2)In view of the problem that the BDD100 k data set does not fully conform to the complexity and diversity of the actual road scenes in China,this paper expands the BDD100 k data set based on the actual road situation in China,and uses K-means algorithm to re-cluster the priority box of the BDD100 k data set and the expanded data set.The detection rate of the model to border regression is improved.Finally,the experimental results show that the expanded data set can effectively improve the recognition accuracy and reduce the recognition time.(3)This paper proposes an improved YOLOv3 signal light identification method for the absence and misdetection of small target traffic signal light recognition.By improving the multi-scale prediction structure,this method establishes the feature fusion target detection layer with the output of quadruple descending sampling,and improves the convolutional layer in front of the target detection layer.By making full use of the details extracted from the low-level feature map,the accuracy of the model for small target traffic signal light recognition is improved.The experimental results show that the improved method proposed in this paper has a good recognition effect in the test set,which is better than YOLOv3,YOLOv4 and Faster-RCNN network.(4)In order to prove the effectiveness and superiority of the proposed signal lamp detection and recognition method,a comparative experiment was carried out in the expanded BDD100 k data set with the traditional machine learning method.The experimental results show that the m AP value is increased by 10.55% compared with the traditional method,and the identification time of a single signal lamp image is reduced by about 33.8ms.(5)On the basis of the research results of this paper,combined with the practical application requirements,design and implement the traffic signal light detection and identification system based on PyQT5 framework.