Research On Adaptive Weight Semi-dense ICP Algorithm

In recent years,with the development of computer vision technology,Point cloud registration algorithm,which is the core technology of 3D reconstruction and visual odometry of SLAM system,has been widely used in virtual reality,artificial intelligence,autonomous vehicles,medical assay and other fields.With the advent of RGB-D sensors like Kinect,ICP algorithm has become the most widely used algorithm in point cloud registration algorithm.How to improve the registration accuracy,registration efficiency and robustness of ICP algorithm has become a research hotspot for most scholars and researchers.In order to solve the problem of large registration error and low registration efficiency in the original ICP algorithm,this paper proposes an semi-dense ICP algorithm based on SIFT feature point neighborhood.The algorithm firstly makes use of the depth information to eliminate the mismatch of feature points,and obtains a more accurate coarse registration result.Then,in the fine alignment,based on the matching SIFT feature points,the algorithm selecting its neighborhood as the matching range,which not only reduces the number of points participating in the registration,but also ensures that the points participating in the registration are located in the overlapping part of the point cloud,and at the same time sets a suitable matching area for each point,which improves the calculation efficiency and reduces the registration error.In addition,the ICP algorithm uses the same fixed weight to calculate the error of the matching point pair in the loop iteration process.The optimal registration result cannot be calculated and the convergence speed is slow.In this paper,according to the curvature information,the Hessian matrix determinant and the cross arm of the matching point pair,the weight is set.Based on the SIFT feature point neighborhood,an adaptive weighted semi-dense ICP algorithm is proposed.Experimental results show that the adaptive weighted semi-dense ICP algorithm has smaller registration error,higher computational efficiency than the ICP algorithm and other improved ICP algorithms.Finally,a SLAM system is implemented using the adaptive weighted semi-dense ICP algorithm as a front-end visual odometer.Experiments on the TUM dataset show that compared to other SLAM systems,the absolute trajectory error of the obtained trajectory results is smaller.