Research On Traffic Light Detection And Recognition Based On Improved YOLOv3

In recent years,the research on assisted driving technology and driverless technology has become more and more in-depth,and traffic light detection and recognition technology,as an important part of environment perception for driverless car,has also received more and more attention.This paper relies on the National Key Research and Development Program(2020YFB1313400),the National Natural Science Foundation of China(U1864204)and the Central University Basic Research Funding Project(300102220204).After in-depth research on domestic and foreign traffic light detection and recognition technology,this paper proposes a traffic light detection and recognition algorithm based on improved YOLOv3.The main research work of this paper includes the following three aspects:(1)An algorithm for extracting regions of interest with traffic lights is proposed.First,based on the difference between the gray histograms of the foggy image and the non-fog image,the foggy day is judged.After comparing the three common defogging algorithms,an improved defogging algorithm is designed to optimize the traffic light image.After analyzing the brightness difference of the traffic light image in the day and night environment,different methods of extraction of the region of interest are selected: in the daytime environment,the black rectangular light board of traffic lights is detected and used as the regions of interest for traffic light;at night,the red ratio and green ratio coefficients proposed in this paper are used to segment the traffic light image in the RGB space,and the red and green suspected traffic lights are extracted as the area of interest.(2)On the basis of preserving the advantages of the YOLOv3 network,and aiming at its insufficient detection and recognition accuracy of small targets such as traffic lights,an improved YOLOv3 network based on traffic lights is proposed.The attention mechanism and cross-channel feature fusion strategy are used.The split-attention module builds a new backbone network to strengthen the new network’s ability to extract features of traffic lights.Secondly,before the new network uses multi-scale fusion feature maps for traffic light detection and recognition,the SE module is added to optimize the obtained feature maps and strengthen the sensitivity of the new network to traffic light features.(3)The traffic light detection and recognition algorithm based on the improved YOLOv3 proposed in this paper was evaluated and analyzed,and the effect was tested on the Bosch traffic light dataset and LISA dataset.Finally,the traffic light images collected by real vehicles in different environments are detected and identified,which verifies the feasibility and robustness of the algorithm proposed in this paper in day,night and foggy environments.