Research On Indoor Vision SLAM Algorithm Based On RGB-D Camera

Visual SLAM(Simultaneous Localization And Mapping)enables mobile robots to locate autonomously and perceive the semantic information of environment in unfamiliar indoor environment,which is the premise for mobile robots to complete intelligent interaction,autonomous navigation and other advanced complex tasks.Visual odometer and image building are two important parts of visual slam.Visual odometer uses the image information of two adjacent frames to estimate the camera's current pose.Image building refers to the process of using pose to generate point cloud map.According to the characteristics of indoor scene,this paper proposes a visual odometer method based on RGB-D image and a method of creating semantic map.This method can effectively improve the positioning accuracy and robustness of visual SLAM algorithm,and realize the semantic awareness of the environment.The main work of this paper is as follows:(1)Indoor scene has the characteristics of low texture and dense repeated structure at the same time.When using the visual odometer method based on image feature points to locate and map,due to the lack of feature points and many mismatches,the positioning accuracy will be low,even the tracking will be lost.In this paper,a visual odometer method based on point,line features and GMS(Grid-based Motion Statistics)is proposed.To solve the problem of insufficient feature points,this method combines point and line features to double the number of feature extraction in low texture and dense repeated structure areas.To solve the problem of many mismatched points,GMS algorithm is used to quickly convert a large number of coarse matching point features into high-quality effective matching points,and RatioTest algorithm is used for fine matching of line features.The experimental results show that compared with ORB-SLAM2 algorithm,the absolute trajectory error of this algorithm is reduced by 11.82%and 14.40%respectively,and the relative pose error is reduced by 3.34%and 7.41%respectively,which effectively improves the positioning accuracy and robustness of the system.(2)The map generated by traditional visual SLAM only contains simple geometric information and location information,does not contain semantic information,can not describe complex environment,and can not meet the needs of mobile robots to complete high-level complex tasks.To solve the above problems,this paper proposes an algorithm of 3D semantic map generation based on YOLO(You Only Look Once)v3.This algorithm uses the YOLOv3 object detection network to detect the key frame image,and fuses with the original point cloud to generate the point cloud with color information as the prior knowledge of clustering.The hyper voxel clustering is used to cluster the point cloud after fusion,and then uses the method of graph cutting to complete the segmentation of the object,and finally establishes the semantic map.Experimental results show that compared with ORB-SLAM2 algorithm,the visual odometry algorithm based on point,line features and GMS proposed in this paper can generate higher quality point cloud data with color information,and more objects can be segmented;Compared with region growing segmentation algorithm,the 3D semantic map generation algorithm based on YOLOv3 proposed in this paper can effectively segment most objects.