A Research Of Objects Detection For Autonomous Vehicle Based On 3D Point Cloud Analysis

Intelligent vehicle technology is the important development direction of future automobile development,which have received more and more attention by the society.In order to realize the safety of intelligent vehicles,the technology of real-time environment perception and dynamic object detection in road scene has become the key point and hot spot in the field of intelligent networked vehicles,which are facing a series of problems to be solved.The technology of environment perception and object detection based on machine vision attracted the highly concern from domestic and foreign intelligent vehicle research and development institutions and industry,due to the advantage of low cost,informative and convenient to use.The related new technology innovation and engineering application develop rapidly.The technology roadmap of fusion Lidar and machine vision will become the mainstream of intelligent vehicle environment perception.In this thesis,the main purpose is 3D mapping for intelligent vehicle and dynamic object detection.A new algorithm of 3D point cloud stitching and point cloud segmenting is proposed to solve the problems of accuracy and real-time in vehicle-mounted camera-based environment perception and object detection,which has obvious theoretical significance and application value.The main contents of this thesis are listed as follows:1.The problem of mapping based on machine vision is studied.For the problem of using binocular camera to acquire images to build an efficient,accurate and real-time 3D map for intelligent vehicle understanding,a stereo images circle matching completeness algorithm was proposed,which considered the accuracy of generated point cloud from corresponding features matching.Firstly,the proposed algorithm selected the feature extraction operator with the highest matching score to extract and match the features of binocular images,so that the corresponding features of stereo images generate point clouds more accurately.Secondly,while building a multi-view map,for the movement of camera,the multi-view scene is constituted.Considering the problem of redundancy of the same name points in overlapping region,an algorithm based on the completeness of stereo image features matching is proposed to reduce the same name points while point cloud stitching.Thirdly,considering the problem of multi-view point cloud registration,a perspective-n-point based local bundle adjustment optimization algorithm was proposed to optimize the 3D points and transformation matrix.Fourthly,considering the influence on point cloud distribution in overlapped region when the same name point is eliminated,an endpoint elimination algorithm based on overlap region center is proposed to reduce the influence of points elimination,so that the point cloud representation of the target contour is not affected.Fifthly,considering the problem of the availability of generated point cloud,a point cloud classification method based on plane fitting is proposed.By this algorithm,the basic plane of the stitched point cloud is removed,and the points outside the basic plane and the corresponding objects are formed as a whole scene,so that it can accurately show the efficient construction of the map.Not only it can quickly express the objects point cloud that the intelligent is concerned about,but also it can roughly judge the volume of the objects point cloud.2.The 3D objects detection of stereo vision based on scene flow and v-disparity estimation is studied.Firstly,for the problem of fast 3D object detection and its real time display in 2D images by combining scene flow and v-disparity estimation,an algorithm combining scene flow and v-disparity estimation to get the proposal position and contour boundary of the objects in the image is proposed.The proposed algorithm can be used to obtain the proposal of stereo image sequence objects in the existing boundary region,and the sparse point cloud of objects can be formed by local reconstruction in the proposal region of corresponding images.Secondly,considering the problem of uncertainty of seeds generating in point cloud segmentation based on region growth,a method of region growing based point cloud segmentation with given sparse point cloud as seed points is proposed.The algorithm can quickly segment the point cloud to get the objects point cloud of specified classification,and the 3D boundary box of the specified object point cloud is obtained.Thirdly,considering the problem of real-time representation of 3D objects in the corresponding images,the 3D object frame represents the object detection results in the corresponding 2D images,an improved 3D bounding box estimation algorithm is proposed.The pose and motion estimation of 3D objects in real scene are realized by this algorithm,and project the 3D bounding box into corresponding 2D image.3.3D object detection based on stereo image edge and stereo pixel estimation is studied.Firstly,aiming at the problem of fast object detection in stereo images,a two-threshold object proposal method based on image edge information and stereo pixel estimation is proposed.Through the fast positioning of the object in the image,the proposed object region can be rapidly generated in the corresponding stereo image sequence,and the 3D point cloud can be reconstructed locally in the object region.Secondly,considering the problem that the edge candidate points caused by multiple targets which cannot determine the target boundary,a judgment rule of distance sum between candidate edge point and adjacent edge point is proposed to distinguish candidate point belonging to the same object or different objects.Thirdly,considering the problem that the candidate edge point is the inner point of the target or is close to the inner point,the boundary of the target is not closed.A judgment rule for the sum of the angles between the candidate edge point and the nearest five candidate points is proposed to determine the candidate point as the edge point.Fourthly,considering the problem of uncertainty of projection of target edge points and stereo pixel estimation points in two-dimensional images,a local bundle adjustment algorithm based on stereo edge characteristics and stereo pixels reprojection error is proposed,and the optimized target edge points are accurately located by this algorithm.Fifthly,considering the problem of detection and location of target point cloud,the distance between the cloud normal vector of the local object point cloud and the cloud normal vector of the whole scene is matched to segment the object.