Road Driving Area Detection Based On Semantic Segmentation

With the rapid development of the economy,the living standards of the people continue to improve,and the number of cars in use continues to grow.The intellectualization of vehicles is receiving more and more widespread attention.The development of autonomous vehicles can alleviate traffic congestion and reduce traffic accidents,which is of great significance in improving the utilization rate of road resources.Among them,perception of the driving environment is a prerequisite for autonomous vehicles to make decisions and control,and enabling vehicles to observe and think like humans has always been the direction of continuous efforts by researchers.With the development and progress of deep learning,the ability of computers to handle images and videos has also improved.The environmental perception of the surrounding environment of vehicles based on visual perception has become a key sense of perception of fully autonomous driving.This thesis uses the identification of the vehicle’s driving area as the research task,and proposes a new semantic segmentation network based on encoder-decoder architecture to achieve accurate and real-time detection of vehicles in a variety of driving scenarios.The main work of the thesis is as follows:(1)Drivable area detection.The thesis first analyzes the problems faced by detecting the driving area and introduces semantic segmentation to complete the identification of the driving area.Secondly,the selection of different feature extraction networks is selected to build a driving area detection network model,and the effectiveness of the cross stage partial(CSP)feature extraction network structure is verified through the model evaluation index comparison and detection effect.(2)Lane line detection within the driving area.In order to better realize the perception of the regional driving area of the structured road,this thesis realizes the detection of the lane line through semantic segmentation,and uses a attention mechanism to handle the feature maps extracted by the backbone network.The upper sampling of the feature maps is used to restore the image size,and the pixels in the image are predicted by the lane line,and the detection of lane line pixel levels is realized.The effectiveness of the newly added attention mechanism module is verified by experimental comparative analysis.(3)Multitask network architecture construction.In order to better achieve the driving area detection effect,the detection network of the driving area and the lane line as different network branches has built the encoder-decoder multi-task network architecture with better accuracy,robustness and real-time,and is not limited by road types.Select CSP-Darknet as the backbone network,designed a branch network that can be detected by the driving area and lane lines,respectively,and build an encoderdecoder semantic segmentation network.After testing and verification,the model can achieve high-precision real-time driving areas and lane line detection in different driving scenarios.