Point cloud matching method and device based on CNN-KNN

Abstract

The invention discloses a point cloud matching method and device based on CNN-KNN, and relates to the field of point cloud matching. The method comprises the following steps: performing feature extraction on two input pictures by using a CNN to obtain respective feature maps; searching the obtained feature points between the two feature maps by using the KNN, and finding out a matching point between the two feature maps; intercepting pixels with specified lengths around the matching points, inputting the pixels into ANN to judge the matching quality, and determining whether the matching points are matched correctly or not. The device comprises a CNN module, a KNN module and an ANN module. According to the invention, the registration precision when the spatial angle is too large can be solved, the registration quality can also be guaranteed, the data-driven model has relatively strong processing capability on large data volume and controllable algorithm time and spatial complexity, and efficient and high-resolution point cloud matching is realized.

Description

technical field [0001] This application relates to the field of point cloud matching, in particular to a method and device for point cloud matching based on CNN-KNN. Background technique [0002] At present, in the fields of computer vision and SLAM, due to the incomplete point cloud data collected by the depth camera and the misalignment of rotation and translation, it is necessary to register the local point cloud in order to obtain a complete 3D point cloud. The points obtained from each viewing angle are collected into a unified coordinate system to form a complete registration of 3D point cloud data. [0003] There are many existing point cloud registration methods, including the histogram method based on the construction of geometric characteristics, the ICP algorithm based on the distance measure, the registration algorithm based on the geometric shape, and the point cloud registration method based on deep learning, etc. . Among them, the histogram method, such as t...

AI Technical Summary

Problems solved by technology

[0004] The above-mentioned algorithms can achieve better results for simple point cloud data registration in the scene, but in some application scenarios, such as SLAM when dealing with inter-frame registration to solve the ring closure problem, the point cloud data obtained from the depth camera, There are often situations where only part of the local point clouds overlap, and the rotation and translation transformation angles are large, resulting in poor matching quality, and it is often difficult to achieve ideal results.

Images