Research On Indoor Visual Slam Algorithm For RGB-D Image

The low cost of depth cameras and the limited range of indoor scenes have greatly promoted the development of indoor visual SLAM based on RGB-D images.Closed-loop detection and scene mapping are important modules for RGB-D SLAM to realize positioning and navigation and maintain global consistency.Closed-loop detection is used to identify places that have been visited;mapping is used to generate point cloud map from observation information.The current RGB-D SLAM has two problems: on the one hand,traditional mapping makes full use of the geometric information of the scene,but ignores the semantic information,which limits the robot's task capability;on the other hand,the commonly used loop closed-loop detection algorithm has two problems: one is that the motion of adjacent frames is used to measure the closed loop,and the cumulative error has a greater impact;the other is that the impact of indoor dynamic points on the accuracy of the system is ignored.In response to the above problems,this paper proposes a closed-loop detection and Octomap improved algorithm based on RGB-D SLAM and an indoor visual semantic SLAM system based on PSPNet respectively.The specific research content of this article is as follows:(1)An improved closed-loop detection and Octomap algorithm based on RGB-D SLAM is designed.The improved closed-loop detection algorithm combines the curvature of the robot trajectory with the loop closure detection algorithm,and the choice of local closed loop or global closed loop is determined by the magnitude of the trajectory curvature.The Octomap improved algorithm combines the two-sided confidence interval of the Gaussian distribution,uses statistical theory to provide a basis for the initial value of the average K-nearest neighbor distance,and then converts the filtered point cloud into Octomap.The experimental results show that after the improvement,the average memory consumption of the point cloud map decreased by about 11.4%,and the average outlier was reduced by11.3%;the average memory consumption of Octomap decreased by about 26.7%,and the average outlier was reduced by 27.3%.(2)A PSPNet-based indoor visual SLAM scene analysis algorithm is proposed.PSPNet uses the pre-trained ResNet network to extract the features of the input image of the SLAM system firstly,and then uses the pyramid pooling model to pool,convolution and upsampling the features of different scales to fuse the multi-scale features,and finally pass through a layer of volume Layered to get the analytical information at the pixel level of the image.The parsed semantic tags are fused and updated in real time on the point cloud using Bayesian update rule.Experiments show that PSPNet has a good analysis effect on indoor scenes and can obtain accurate semantic tags.(3)A three-dimensional semantic map generation algorithm based on scene analysis is designed.Using the epipolar constraint model,the dynamic points of the scene can be identified and eliminated by solving the distance from the projected coordinate point to the epipolar line.After removing dynamic points,the global semantic map of the system is constructed by combining the real-time semantic information obtained from the scene parsing network.Experiments show that the system can construct a global consistent 3D semantic map accurately.Compared with ORB-SLAM2,the front-end performance improves by about 5% in absolute trajectory error and 8% in relative trajectory error on average.