Research On Accelerate Methods For Evolutionary Point Cloud Registration Technique

Three-dimensional point cloud data has been an important data form in computer vision area,along with the rapid development of the depth-acquisition device.It is difficult to obtain all the point cloud information due to the restriction of the field of view and the occlusion of the shape.Multiple views of point clouds need to be collected and then aligned to a uniform coordinate to depict the object completely.Therefore,accurate,efficient and robust point cloud registration method is a hot topic in three-dimensional imaging.Evolutionary point cloud registration method is a kind of registration method,which is innovative,high-precision and robust to initial positions of point clouds.However,this kind of method usually time-consuming,which effects the application on engineering area.Thus,this paper focuses on the accelerate strategy of the evolutionary point cloud registration method and carries out the following study:1.Many state-of-the-art registration methods,including various rough registration methods,classical ICP fine registration methods and two variant,evolutionary point cloud registration methods,are discussed and analyzed.Their applicable conditions,limitations and improvement directions are also pointed out.2.To reduce the over-exploitation caused by the resolution of point cloud is lower than the convergence precision in evolutionary point cloud registration methods.An evolutionary point cloud registration method based on hash table and moth-flame optimization is proposed.The method uses a hash table to quickly insert and look up object function values to avoid over-exploitation and repeated search in individuals' updating.Additional,a new search equation and restart mechanism is involved to balance the exploitation and exploration ability of the method.Finally,an improved moth-flame optimization algorithm is utilized to find the optimal Euclidean transformation.Compared with some state-of-the-art evolutionary point cloud registration methods.The experiment results show that our methods can efficiently reduce the processing time.3.Traditional evolutionary point cloud registration methods often not using the color information in the models.To overcome the defect,this paper introduces a point cloud registration method based on self-adaptive evolutionary optimization algorithm and color information.The input point clouds are subsampled by extracting the color feature points and randomly chosen points,the median of all pairs of color constrained points is utilized as the object function.At last,the self-adaptive evolutionary optimization algorithm is used to get optimal solution.The registration experiments on four colorized point clouds show that,compared with the evolutionary point cloud registration methods only spatial information is used in and two state-of-the-art registration methods,our method significantly shorten the processing time while achieving similar registration precision.