Research On Detection Method Of Lane Line And Traffic Sign In Complex Scene

Lane line and traffic sign detection technology,as the core technology of the sensing system in autonomous driving technology,can not only ensure the safe driving of vehicles,but also improve the driving efficiency of vehicles,reduce traffic congestion and traffic accidents.With the rapid development of artificial intelligence technology,a large number of target detection and recognition algorithms continue to emerge,but due to the complexity and variability of the environment in which lane lines and traffic signs are located,existing lane line and traffic signs detection methods have resulted,but have not taken into account accuracy,robustness,and real-time at the same time.Therefore,lane line and traffic sign detection technology have been hot research topics in computer vision and autonomous driving.This article studies lane and traffic sign detection in complex environments.The main contents are:(1)To solve the problems of vehicle occlusion and ground fouling,The thesis considers lane line detection as a segmentation problem of continuous and slender regions,and proposes a lane line detection method based on a dense segmentation network.First,based on the sementic segmentation network(Seg Net),a dense segmentation network is constructed using dense blocks.The constructed deep segmentation network establishes a connection relationship between different layers,making the network have the characteristics of reusing features,learning high-level instance features,and reducing the loss of spatial information as the network deepens.Influence,improve the accuracy of image segmentation;secondly,in order to make the lane line boundary more specific,the method introduces the neighboring AND operation and Meanshift clustering algorithm to process the output of the dense segmentation network to reduce the interference of non-lane line pixels.Finally,in order to be able to perform segmentation of lane lines,thesis makes a segmentation dataset of lane line instances corresponding to the Tu Simple training set,and trains the network end-to-end together with the Tu Simple training set.Experiments show that the algorithm proposed in thesis can solve the problems of vehicle occlusion and ground fouling,has good robustness.(2)To solve the problem of low efficiency of traditional convolutional neural network,thesis proposes a traffic sign detection method based Thunder Net to detection traffic signs in real-time in complex scenes.Thunder Net is composed ofbackbone and detection.In order to use more context information and encode a larger range of adjacent pixels to improve the effect of target detection,the backbone part is based on the Shuffle Net V2 network model,using depth separable convolution to replace the depth separable convolution to expand the receptive field The constructed network model.The detection part draws on the Light-Head R-CNN network model,and also uses a context enhancement module(CEM)to aggregate multi-scale global information and local information to generate more differentiated features and improve features Representation capabilities of graphs.Spatial attention module(SAM)is also used to optimize the feature distribution to emphasize traffic sign information and weaken the background information.At the same time,the region proposal network(RPN)is compressed to improve efficiency.Experiments show that the method proposed in thesis not only has good accuracy and robustness,but also has good real-time performance,and the detection speed can reach 120 fps.