Research On Low-overlap Point Cloud Map Registration

With the development of industrialization and the arrival of the information age,intelligent robots have played an important role in manufacturing,transportation,Internet of Things,and intelligent services.Indoor intelligent robots have carried out research and development on autonomous intelligent cleaning platform technology for cleaning tasks.First of all,it is necessary to realize the collaborative construction of high-precision semantic maps of the scene and the independent incremental update to ensure that they can accurately find their own position and avoid obstacles in complex environments,and then complete the cleaning task.For the working environment,because the work is done indoors,the GPS(Global Positioning System)signal is poor and the positioning accuracy is greatly reduced,so the three-dimensional mapping based on lidar has become a research hotspot.After reading a lot of related literature,it is found that most of the point cloud registration methods in existing mapping are aimed at scen es with high overlap rates.For environments with more occlusions and fewer corresponding points,such as in narrow passages,corners,etc.,there is not much research on the registration of the two frames of low overlap point cloud.In order to apply point cloud registration technology in complex scenes more widely and flexibly to intelligent robot scene recognition and provide strong technical support for obstacle avoidance,also provide help for the development and promotion of cleaning robots and their real practicality.Aiming at the above problems,this paper has gone into and proposed reasonable solutions to achieve accurate and efficient image registration algorithms in low overlap regions.The main contents of this article are as follows:1.Propose an Iterative Closest Points(ICP)algorithm based on improved Point Pair Features.Filter the effective key points by limiting the angle between the normal vector of the point and the point,and eliminate the noise generated by moving objects/people.Then through clustering,boundary point elimination,and neighbor point screening operation,the key points are further extracted,point cloud is simplified,and the overlap of the region to be registered is improved,and the point-to-plane error function is used as the registration error.On the one hand,this algorithm overcomes the disadvantage that traditional ICP cannot achieve point cloud map registration in low overlap regions.Experiments show that this algorithm has good registration accuracy.On the other hand,based on this algorithm,it can speed up 3D modeling by reducing the number of sampling frames,which greatly saves time.2.Adopt an ICP algorithm based on HMRF(Hidden Markov Random Field).Based on improving the key points of point-to-features,starting from the aspect of non-corresponding point elimination,HMRF modeling is proposed,using its Markov property to establish the relationship between points and the points in the domain system.And through EM(Expectation-maximization)algorithm infers whether a point is the correct corresponding point.Experiments verify that this algorithm can further improve the accuracy of low-overlap point cloud map registration and remove incorrect matching relationships.