| With the continuous improvement of equipment performance and the gradual popularization of three-dimensional sensors,computer vision tasks are gradually developing in three dimensions.Lidar is widely used in military and civilian applications at home and abroad because it can provide more accurate positioning information,better characterizing shape,and strong anti-interference ability.The unstructured point cloud data acquired by lidar is characterized by large quantity,disorder and sparsity,as a key technology of 3D environment perception,it is very important to study the 3D object detection and recognition algorithm based on lidar point cloud data.In the road environment,Using lidar sensor as the environment sensing method,this thesis focuses on the three-dimensional ground object detection and recognition algorithm based on lidar,including multi-stage and single-stage schemes: multi-stage scheme research is divided into ground point filtering,object clustering,object recognition,and corresponding algorithms are proposed respectively;single-stage scheme research is based on the time-consuming problem of multi-stage scheme,and a Point Yolo object detection and recognition network model based on deep learning is proposed.The research content of this thesis is as follows:1)In the preprocessing research of lidar point cloud data,the related indicators of Velodyne-HDL-64 E lidar are briefly introduced;the characteristics of horizontal and vertical distance resolution of the lidar are analyzed;at the same time,the black and white point noise in the process of point cloud data acquisition are analyzed,and the neighborhood interpolation completion algorithm and the KNN(K-Nearest Neighbor)numerical statistical filtering algorithm are used to complete the black point and filter white point respectively.The feasibility of filtering out the noise by the above algorithm is proved by experiments.2)In the multi-stage scheme research,for the ground point filtering,a ground point filtering algorithm based on front view is proposed to solve the problem of excessive filtering,while avoiding the influence of ground points on ground object segmentation;for object clustering,a hierarchical Euclidean distance clustering algorithm is proposed to solve the problem of distance-similar object clustering;for object recognition,the typical characteristics of customization are proposed,and datasets are created to train MLP(MultiLayer Perception)fully connected network models.Experiments show that the above algorithm can effectively complete the object detection and recognition of multi-stage scheme.3)In the single-stage program research,the 3D point cloud dataset is introduced,and the construction of the Point Yolo model is illustrated according to the characteristics of the point cloud data.Experiments show that the Point Yolo model can effectively complete object detection and recognition,predict the bounding boxes and categories of objects at the same time,and effectively solve time-consuming problems compared with multi-stage scheme. |
