Research On Road Detection Based On LiDAR And Visual Image Data Fusion

With the increasing economic development,the position of unmanned driving in automobile driving is becoming more and more important.In the unmanned driving,the key part of the vehicle’s various driving decisions is road detection.However,the existing multi-modal data fusion method applied to road detection fails to reduce the distance between LiDAR point cloud data,which leads to the problem of high cost of detection process time in road detection.In this paper,two heterogeneous data of LiDAR point cloud data and visual image data are fused at the data level through spherical coordinate transformation,and the improved SegNet is used for road detection,so as to achieve a high-precision and low-latency road detection.The main research contents are as follows:(1)Road detection based on LiDAR point cloud and full convolutional neural network.First,the LiDAR unstructured point cloud data is processed to generate a top view image with encoded feature data(such as average height and density).Through the top view image,road detection can be realized based on visual image data,which is implemented by a simple and fast Full Convolutional Neural Network(FCN).(2)Road detection based on visual images and SegNet.First,build the SegNet network framework for segmenting the passable area of the road,which is an efficient architecture for pixel-level semantic segmentation.In the SegNet,stochastic gradient descent(SGD)is used to optimize the weights in the network,thereby increasing the end-to-end training speed in the network,and speeding up the use of visual images for road detection.(3)Road detection based on heterogeneous data fusion based on improved SegNet.First,the visual image data and point LiDAR cloud data are preprocessed.In this paper,spherical coordinate transformation is introduced to reduce the gap between the 3D LiDAR data point clouds and optimize the calculation time.The visual data of the camera and the conversion data of the lidar point cloud are input to the improved SegNet with extended receptive field for road detection.Experimental results show that the LiDAR point cloud-visual image fusion method proposed in this paper is faster than other current fusion methods while maintaining accuracy.