Research On Obstacle Detection Algorithm For Intelligent Vehicle Based On LiDAR And Vision

Obstacle detection is a challenging research task in the perception layer of advanced assisted driving.As the basis for intelligent vehicle(Ⅳ)decision-making and execution,it is a prerequisite for safe driving.At the same time,obstacle detection technology is an important guarantee for the safe driving of IV.Light Detection and Ranging(LiDAR)has the advantages of high accuracy,strong stability,and resistance to weather.Not only it can obtain the position and three-dimensional information of obstacle in complex environments,but also it is widely used in the field of object detection.However,due to the sparseness of the point cloud,it is easy to cause incomplete obstacle information.Vision cameras can form a good complementary relationship with LiDAR,which obtain complete image information of obstacles.Also,it has the advantages of low cost and high resolution.Therefore,how to perform the fusion of LiDAR and camera information to improve the accuracy of object detection is the main research content of this paper.Firstly,the data modalities of LiDAR and camera sensors,the commonly used object feature extraction and classification methods are introduced.At the same time the multi-sensor information fusion scheme is analyzed and summarized.The data sets and evaluation indicators used in object detection are introduced in detail.After that,the selected LiDAR and camera are described.The common coordinate system of IV is established and the mathematical model of the camera is constructed.The calibration of LiDAR and camera parameters based on the calibration board is completed and the existing problems are analyzed.Then,in the traditional joint calibration of LiDAR and camera based on deep learning,the input of the network is the original point cloud and original image data.Feature extraction produces redundancy,which leads to slow network inference.To solve this problem,a depth map-based method is proposed.The LiDAR and camera joint calibration algorithm can effectively reduce the color and other features that are useless for calibration.The algorithm converts the visual picture into a depth map,and the point cloud with error obtains its depth map in the image coordinate system,as the input of the deep learning network.The number of features is reduced.Also,the depth and width of the network are reduced.So that the processing speed of the deep learning network is accelerated,thereby reducing the training time and improving the real-time performance of the system.The validity of the joint calibration algorithm of LiDAR and camera based on depth map is verified by the dataset and measured data.Finally,aiming at the problem that the point cloud information of the obstacle target is incomplete and the point cloud of the occluded object is scarce,which leads to the decrease of detection accuracy,a joint target detection algorithm based on point cloud enhancement is proposed.The algorithm feeds back the center coordinates predicted by the first Region Proposal Network(RPN)back to the point cloud enhancement module,which provides the obstacles imformation for point cloud enhancement module.Symmetric point clouds are acquired by symmetry.However,the point cloud in the area is enhanced by using spherical sampling,which introduces too many false point clouds.A cylindrical sampling method is proposed to enhance the accuracy of the point cloud and improve the accuracy of the object detection algorithm,especially for obstructed object.Through the experience of the dataset,proposal scheme is superior to the original algorithm in terms of training speed and accuracy,which verifies the effectiveness of the point cloud enhancement algorithm.