Design And Implementation Of 3D Pointcloud Scene Recognition Based On Bag-of-words

The cognition of self-position is very important for robotics,as it is the foundation of any action tasks.For the robot in complex working environment,the correct recognition of the scene is not only conducive for the recognition and perception of the key targets in the environment,but also very effective to assist the positioning module.In addition,scene recognition technology has been already widely used in many loop closure methods,as the part of Simultaneous Localization and Mapping(SLAM),which is a core field of robotics.Odometry-based SLAM method's essential idea is updating the map iteratively.This operation will inevitably introduce the accumulated error,which is especially obvious in large-scale mapping.Loop closure uses scene recognition technology,which can detect the historical scene when the robot's motion trajectory is looped,recognize the accumulated error and fix the map accordingly,so as to reduce the accumulated error.In this study,a new robust scene recognition method based on the bag-of-words in stereo pointcloud is designed and implemented for the application scenario of autonomous driving.Based on this method,the functional tests of loop detection have been carried out on the open dataset.This study attempts to directly apply the bag-ofwords technology to process raw stereo pointcloud,and proposes the description based on overall features according to the characteristics of stereo pointcloud data.For geometric consistency checking,based on Random Sample Consensus(RANSAC),this study implements a method of filtering homonymous point pairs based on adjacent clusters,which combines spatial information and local description to filter homonymous point pairs.Also,it proposes a quadratic geometric verification method based on probability density distribution,which greatly improves the accuracy of geometric consistency checking.Experiments show that the implemented bag-of-word-based scene recognition method in stereo pointcloud can effectively work in the data scene of autonomous driving,and realize the loop detection ability with high robustness.The application of bag-of-words-based loop detection can match the correct historical frames under the condition of only LiDAR data without odometry,and assist the SLAM to realize loop closure,which is the basis of large-scale SLAM.