Research And Implementation Of Landmark Model Map Based On Monocular Robot

With the development of unmanned driving and robotic technology,machine vision has been widely used in the field of autonomous navigation.Visual SLAM extracts and matches the observed data through visual sensors,and finally generates a map containing the information of the surrounding environment.The location and navigation of the robot depend on the map data.Monocular vision,as a branch of visual SLAM,attracts much attention because of its low cost and easy access to sensors.However,compared with other sensing methods such as laser,binocular and RGBD,it is difficult for monocular cameras to directly recover environmental information from observations due to their observation methods.In the traditional monocular vision SLAM research,the main problems of map construction are as follows: 1)Sparse point cloud can be positioned and can not describe entities;2)Dense point cloud map data is too large;3)Point cloud-based map can not support navigation and obstacle avoidance very well.Aiming at these problems,this thesis proposes a topological-landmark map based on landmark model reconstruction,and improves the classical mapping algorithm to make it possible to reconstruct landmark model.In order to improve the maps built by existing monocular vision schemes,this thesis mainly studies the following aspects:(1)In order to improve the performance of semantic map in monocular vision robot location and navigation,,this thesis proposes a multilevel map and mapping method based on landmark model,which combines semantic segmentation,point cloud data,landmark model index and regional topological map’s data structure.(2)Aiming at the problems of existing algorithms in generating semantic maps,this thesis proposes a method of constructing maps by using landmark model for semantic reconstruction,and proposes a feature extraction algorithm of landmark model combined with existing semantic segmentation and depth estimation algorithms.The experimental results show that the method can effectively segment the target entity from the observed data.(3)Aiming at the problem that the sparse point cloud can’t recover the entity and the amount of dense point cloud’s data is too large,this thesis studies the method of using the feature of signpost entity to locate.Combining with the random forest and probability model,this thesis proposes the probability distribution model of robot’s location by observing the signpost entity.Through experiments,the location accuracy and data sparsity are compared.The success rate of relocation and match speed on the open data set KITTI better than BOW,and the data is reduced by about 40% compared with LSD-SLAM’s point cloud.(4)At last,combining region cut and topological structure of the map,the topologicallandmark map structure is constructed through the landmark features,and the analyze the performance of robot location and navigation.The location and navigation experiments are carried out on our robot using ROS platform,and the average location error is 0.35 m.