Research On Vision-Based Detection Method For Unmanned Vehicles In Campus

With the intelligent development of modern transportation,the application of unmanned driving technology has become increasingly widespread.Accurate perception of the surrounding environment is the basis for safe driving of unmanned vehicles.Vision-based passable area detection is one of the important research directions of environment perception technology.Aiming at the environmental characteristics of the campus scene,the research on the detection method about the passable area of unmanned vehicle in the campus based on vision is carried out.The research content is as follows:Firstly,in view of the large differences between the lane markings in the campus,urban traffic and expressways,a lane line detection method based on improved Lane Net and Convolutional Neural Networks(CNN)is proposed because of the poor applicability of existing algorithms.The detection accuracy is improved by optimizing the Lane Net binary segmentation branch loss function,and the detection speed is improved by simplifying the H-Net network adaptive perspective transformation matrix.Based on the lane line segmentation example,the common lane lines in the campus scene are classified through CNN.The experimental results show that the proposed method improves the lane line detection accuracy of the Lane Net algorithm by 1.2% and 1.9% on the Tu Simple and CULane datasets,and has a better recognition effect for common lane line types in the park.Secondly,considering the problems of low obstacle detection accuracy and insufficient real-time performance in the campus scene,an obstacle detection method combining improved YOLOv3 and stereo vision is proposed.The forward inference time is reduced and the model detection speed is improved by optimizing the feature extraction network;then the model detection accuracy is increased by increasing the feature map scale,improving the loss function,and re-clustering the prior frame size;finally,the improved model is applied to a binocular stereo camera to obtain distance information between obstacles and unmanned vehicles.Experimental results show that compared with the YOLOv3 algorithm,the proposed method increases the frame rate by 8fps,m AP increases by 4.19%,and the average distance measurement error of obstacles in the park is as low as 4.67%.Finally,by analyzing the detection results of lane lines and obstacles,a passable area detection method is proposed based on the Robot Operating System(ROS)framework.The lane line detection and classification results and the obstacle recognition and ranging results are published as the ROS topic;then the topic can be subscribed by the passable area detection function package.By delimiting the vehicle driving plane and expanding the obstacles,the passable area topic can be published and visualized.The experimental results prove that the passable area of unmanned vehicles in the campus can be detected accurately.