Road Segmentation Based On Deep Convolutional Neural Network

Autonomous driving is an important part of Chinese next generation artificial intelligence development plan and is also an important area for the development of all countries in the world.However,the current perception technology cannot work well under complex and variable traffic scenarios and weather conditions.Therefore,it is of great significance to construct a perception model with strong generalization ability and high precision performance.Based on the theory of deep convolutional networks,the thesis studies the lane line detection and road segmentation in traffic scenarios.The specific research content is as follows:A vision-based road segmentation algorithm is studied.The imbalance between the road and the background area in the picture leads to slow convergence of the algorithm,the thesis optimizes the loss function of the model.In order to void leak detection and edge misdetection in road segmentation of visual data,the thesis introduces image depth map and constructs a road segmentation model,which includes the feature extraction module of the cavity convolution residual network,feature fusion and cavity pyramid pooling.This thesis also establishes a data volume of tens of thousands of road data sets.Based on self-built datasets and KITTI datasets,the results show that it improves the convergence speed and accuracy of the model effectively.A vision-based lane line detection algorithm is studied.Aiming at the problem that the narrow line of the lane line is easy to cause the lane line breakage problem,the thesis constructs a feature extraction framework based on cavity convolution and a spatial convolution network of the back-end fusion like Markov random field,which effectively mines the correlation features of long-distance pixel points.Reduced lane line classification errors.A road extraction algorithm based on LiDAR data is studied.The thesis uses the visual data to label LiDAR data based on sensor spatial relationships.Aiming at the problem of road edge misdetection,this thesis designs a point cloud segmentation network that combines the features of point cloud normal vector and reflection intensity.The results show that it improves the convergence speed and final precision of the model effectively.Finally,this thesis proposes a road segmentation network that combines two kinds of data.Based on the research of road segmentation of visual and LiDAR data,the fusion model introduces the point cloud network as well as image network and combines two data features to fully exploit the advantages of each sensor data,which achieves an endto-end road segmentation model.Based on KITTI datasets,The robustness and test accuracy of the fusion model are higher than those of the single data road detection algorithm.