Research On Target Recognition Method Based On Three-Dimensional Laser Point Cloud

Three-dimensional objects recognition in a scene is a major challenge in the field of three-dimensional computer vision.In the past,the main research on object recognition was based on two-dimensional data,and it is a relatively mature field at present.Therefore,the research direction has shifted from two-dimensional data to three-dimensional.With the rapid development of three-dimensional sensors,three-dimensional object recognition has been became a research hotspot in recent years.The point cloud object recognition technologies can make use of the environmental information in the scene,but the point cloud data is characterized by noise interference,occlusion and a large number of data,which make the object recognition based on the point cloud has low computing efficiency and low recognition accuracy.To solve these problems,this paper studies the object recognition method based on three-dimensional laser point cloud.This paper,according to the characteristics of point cloud data,focus on studying the preprocessing technologies of three-dimensional point cloud data,key point extraction algorithm,point cloud feature extraction and matching,and target recognition algorithm based on three-dimensional laser point cloud.The performance of the algorithm is verified through data set simulation and real experiment.The main research contents of this paper are as follows:1.An adaptive radius filter is proposed,which solves the problem that the traditional radius filter is not easy to determine the radius.The method needs to first find the volume of the smallest bounding box of the point cloud in space and the total number of points of the point cloud,and then use the cube root of the ratio of the volume to the total number of points as the search radius.Experiments have proved that using this algorithm for point cloud filtering is fast and clear.It does not need to manually adjust the search radius according to the input point cloud,which improves the filtering efficiency.2.The traditional voxelized grid filtering is improved,which solves the problem of the traditional voxelized grid filter that the unit voxel side length needs to be manually determined and the number of points after filtering is uncontrollable.Before running this method,the point cloud is firstly filtered by radius filter to filter out outliers;then determines the down-sampling rate according to the number of points needed,and automatically determines the side length of the unit voxel and the number of points after down-sampling according to the down-sampling rate,which improves the ability to streamline the point cloud.3.A feature matching algorithm based on mutual nearest neighbor which satisfies the specified threshold condition is proposed.The algorithm improves the problem that the mismatch rate of matching points obtained by the current feature matching algorithm is too high.This algorithm combines two feature matching algorithms,the nearest neighbor method and the threshold method,and then searches for the nearest neighbor mutually whose distance is less than the threshold between the model point cloud and the scene point cloud to achieve feature matching.Experiments show that the feature matching algorithm plays a certain role in improving the rate of correct matching points.