Unstructured Road Recognition Based On Image And Lidar Point Cloud Data Fusion

As the ultimate development direction of intelligent vehicle,the safety of driverless vehicle depends on the accurate understanding of the surrounding environment.In the real environment,there are a large number of unstructured road regions with no obvious road features.Area segmentation based on pixel level of passable area can avoid the limitations brought by model assumptions.It is of great significance for road identification,especially unstructured road identification,to divide passable areas by pixel level semantic segmentation of acquired data,which can provide accurate passable areas for smart cars and improve the safety of smart cars.This paper proposes an unstructured road recognition algorithm based on deep learning and data fusion.The main purpose is to extract the passable range of the road.The specific process is as follows:(1)the KITTI data set was preprocessed to obtain images and labels for training,and the point cloud was screened for redundant points and projected in space.(2)the fcn-8s model was established,and the processed data training model was used to obtain the deep learning model for segmentation and conduct segmentation and feature extraction.(3)the processed point cloud was used to segment the point cloud on the road according to the height information and the horizontal information,and then the point cloud boundary was extracted by the alphaShapes algorithm,and the road segmentation was achieved by judging whether the pixel was in the boundary.(4)using the data obtained from the deep learning and point cloud segmentation results,build the MRF model,and use the ICM algorithm to solve the global minimum of energy function,the minimum energy state is the final segmentation result of the corresponding pixel,and by using the industry commonly used indicator of deep learning segmentation results,MRF segmentation results were evaluated and compared.