Semi-dense ICP Algorithm Based On Surf Dimension Reduction Description

Image point cloud registration is widely used and can be used in pattern recognition,motion tracking,image cutting,stereo matching and other application fields.It is an important research topic in computer vision and image processing.The ICP algorithm is the most widely used algorithm in the field of point cloud registration research.In this paper,we will study the matching area and matching residuals in the ICP algorithm.The ICP algorithm changes the transformation relationship between the two images by looping to minimize the matching residual.The construction of matching residual function directly affects the accuracy of matching.The original ICP algorithm used the three-dimensional coordinate Euclidean distance of pixels as the matching residual.There is an error in the depth information of the point cloud image,which may cause the iteration of the ICP to fail to converge to the global optimum and affect the matching accuracy.In order to reduce the impact of pixel depth information inaccuracy on the ICP algorithm,this paper combines the pixel SURF description and spatial three-dimensional coordinate construction to match the residuals.First,the 64-dimensional SURF description is reduced to 4 dimensions,and then all the pixels participating in the registration are subjected to 4-dimensional SURF description.The 3-dimensional coordinates of the pixels are combined with the 4-dimensional SURF description to make the 7-dimensional Euclidean distance as the matching residual..The new matching residuals not only increase the pixel image feature information,but also reduce the impact of depth information inaccuracy on the matching accuracy;and because the SURF description has the characteristics of scale invariance,rotation invariance,and illumination invariance,the ICP has a high Robustness.The image area of ICP algorithm involved in matching will also affect the matching accuracy.The ICP algorithm in which all pixels participate in the matching will have a situation where some pixels in the reference point cloud have no matching pixels in the registration point cloud,resulting in an ICP loop iteration that cannot be globally optimized.Therefore,the dense ICP algorithm is affected by how much the images overlap.This paper proposes a semi-dense ICP algorithm based on SURF feature points.The algorithm selects the spherical neighborhood centered on the SURF feature point in the image as the reference point cloud,and then defines the registration point cloud area in the image to be registered according to the spherical neighborhood.Limiting the reference point cloud and registration point cloud area by the spherical neighborhood of the image feature points can not only ensure that the reference point cloud has a matching point in the registration point cloud,but also can remove some of the depth difference and depth value from the feature point.For inaccurate pixels,the registration accuracy of ICP algorithm is proposed.Finally,based on the SURF dimensionality reduction description,this paper completes a semi-dense ICP matching algorithm and applies it to the SLAM system to provide a high precision front end for the SLAM system.The test of SLAM system in TUM data set has small trajectory error and is practical.