Study On Lane-level Traffic Light Recognition System Of Autonomous Vehicles

The traffic light recognition system of intelligent vehicles is a key part of the automatic driving environment perception function.At present,there are three main challenges in the research of traffic light recognition system:(1)The content of the scene is complex and the environmental conditions are changeable,resulting in reduced signal recognition accuracy and robustness;(2)The semantic information recognition of the traffic light is lacking,and the indication is not strong;(3)Recognition accuracy and realtime are difficult to balance and cannot be adapted engineering applications.In order to solve the above problems,the paper proposes a traffic light recognition system based on deep learning detection,integrated feature channel tracking and lane-level high-precision positioning;built a mobile sensor module based on an embedded computing platform,and installed it for testing.The research content of the thesis mainly consists of two parts:(1)Research on the image recognition algorithm of traffic lights.First,design a deep convolutional neural network model and integrated feature channel tracking method;then establish a mechanism for simultaneous detection and tracking,and run the video reading module,deep convolutional neural network model,and integrated feature channel tracking module at the same time;set the inter-frame targets buffer to filter out target mutations caused by false detection or detection flash;finally,the effectiveness of the proposed method is verified by comparison with existing methods.(2)Research on the traffic light recognition system for determining vehicle-related traffic lights and their semantic information in conjunction with high-precision maps.First,a movable sensor module based on the embedded computing platform is built.In the preparation stage,point cloud maps are made,vector information is manually labeled,and the camera and lidar are jointly calibrated.In the operation stage,the driving lane and the matching traffic light position are determined by lidar and normal distribution transformation algorithm;the aforementioned image recognition algorithm is used to determine the specific position and color information of the signal light from the image content in the candidate area.The research results of the paper show that: the paper reasonably designs the structure of the deep convolutional neural network and the composition of the integrated feature channel,and the composite mechanism of image recognition that combines detection and tracking information is constructed,which is suitable for the image recognition of traffic signals.Compared with traditional model-based methods or single deep learning algorithms,the image recognition algorithm proposed in the paper has better recognition accuracy and recognition speed for traffic lights.The paper uses multisensor data to establish a traffic light recognition system based on lane-level highprecision positioning and image recognition algorithms.The system can accurately determine the traffic light and its semantic information of the lane where the vehicle is located.In addition,a mobile sensor module based on an embedded computing platform was built,and a real vehicle test was carried out to meet the requirements of the traffic light recognition system.