Research On Adaptive Target Extraction Algorithm Of TOF Vector Point Cloud Based On Unsupervised Learning

With the rapid development of artificial intelligence technology,three dimensional computer vision has attracted more and more scholars’attention.Time of flight（TOF）3D imaging is an important research direction in the field of computer vision.Due to the advantages of high frame rate,fast response speed and strong anti-interference ability,it is widely used in unmanned deep space exploration,robot navigation,autopilot,simultaneous localization and mapping（SLAM）and other fields.However,TOF imaging technology will be affected by system noise and non system noise,which brings difficulties and challenges to the three-dimensional scene segmentation and target extraction of TOF point cloud.This paper proposes an adaptive extraction algorithm of TOF vector point cloud based on unsupervised learning to solve these problems.The research content of this paper includes the following aspects:1.The composition of TOF 3D imaging system,imaging physical principle and mathematical model,depth measurement error analysis theory are studied.The basic principles of k-means clustering algorithm,K-NN（K-Nearest Neighbor）classification algorithm,K-NN search algorithm and spatial organization index are studied.The data structure,algorithm flow and time complexity of kd-tree（K-dimensional Tree）index are analyzed in detail.2.The difficulty of adaptive target extraction is caused by the difference of TOF imaging results in different scenes,range modes and resolutions.Combined with the idea of machine learning,an adaptive target extraction method of TOF vector point cloud is proposed,and the applicability of complex target extraction algorithms under different resolutions,different targets,different viewing angles,different distance modes and different shooting environments is studied.The resolution includes 320×240 and 640×480 resolution,the target includes different targets and different numbers of targets,the angle of view includes different shooting angles and different target angles,the distance mode includes 1m,3m and 4m distance modes,and the shooting environment mainly includes experimental platform,indoor scene and outdoor environment.3.The theoretical research of radius filtering,K-NN filtering and SOR（Statistical Outlier Removal）filtering is carried out,and the methods proposed in this paper are compared with them.And this paper studies and compares the processes of linear scanning method and kd-tree index method in dynamic K-NN search.A dynamic K-NN fast search method based on kd-tree local reuse feature cache is proposed.4.The algorithm proposed in this paper is verified and analyzed in detail.Experiments show excellent performance and stronger robustness of bisection iterative k-means algorithm compared with bisection method and clustering method.The target extraction algorithm based on unsupervised learning of point cloud vector has good adaptive extraction effect.The values of F1 and Fincrease by about 30%and 10%respectively compared with V-RS（Vector-RANSAC）method,and about 50%and 30%compared with other traditional methods.Compared with kd-tree method,the dynamic K-NN fast search algorithm based on kd-tree local reuse feature cache improves the computational efficiency by about 51%without reducing the filtering effect.The data analysis of the above algorithms is carried out.The characteristics of each algorithm,the adaptation interval and the influence of the selection of parameters on the algorithm results are summarized.The main innovations of this paper are as follows:1.An adaptive extraction algorithm of TOF vector point cloud based on unsupervised learning is proposed.Step one is to complete the global target location by FVP-k-means（Full Vector Projection k-means）algorithm.Step two is to filter the local noise in the area of the target by the iterative K-NN filtering algorithm based on Gaussian distribution.2.A dynamic K-NN fast search method based on kd-tree local reuse feature cache is proposed.The concept of dynamic K-NN search problem is defined,and the principle of redundant calculation and optimization space are analyzed.The"nearest neighbor table"data structure is used to store the nearest neighbor information of nodes,which can reduce the construction times of kd-tree index and improve the computational efficiency.