Research On Visual SLAM System Based On Semantic Segmentation In Indoor Dynamic Environment

Robust visual simultaneous localization and mapping(SLAM)system in indoor dynamic environment is of great significance to blind navigation algorithms,but the current mainstream visual system only assumes that in a static environment,once the scene has dynamic objects,it will lead to large accuracy The magnitude is reduced.In this paper,based on deep learning,combined with optical flow or geometric methods,a robust visual SLAM system is established,which enables higher trajectory accuracy compared with other systems in indoor dynamic environments.First,the five essential links of the classic visual SLAM system are briefly introduced;the performance characteristics of common open source visual SLAM systems are compared and analyzed,and the appropriate ORB-SLAM2 system is selected as the next research based on the characteristics of the indoor dynamic environment studied in this paper.It elaborates the relevant theoretical knowledge in the selected ORB-SLAM2 system,including ORB feature extraction,feature point matching and camera pose estimation,etc.Three types of input images captured by different camera types are also introduced,and the TUM RGB-D image dataset is selected from the practicality,information content and trajectory accuracy;finally,the performance test analysis of ORB-SLAM2 in indoor dynamic environment is carried out.Secondly,the basic theoretical knowledge of convolutional neural network is introduced in detail;then the network structure of Mask R-CNN and the neural network module Piontrend are introduced,and the two are combined into a semantic segmentation network M+P;Improvement of ORB-SLAM2 system based on M+P semantic segmentation network,and the prior dynamic objects are eliminated before running other threads.Accuracy improvements were found in the dataset compared to the original system.Finally,the experiments of the visual SLAM system based only on deep learning show that the method of only culling the prior dynamic objects,the obtained pose accuracy is average,which shows that there is still a large room for improvement.One of the important ways to improve is detection and rejection of posterior dynamic objects.The principles and steps of LK optical flow method and multi-view geometry are explained in detail,and combines them with the M+P algorithm to test under the TUM RBG-D data set,compare and analyze the test results of the two.Among them,the visual SLAM system based on semantic segmentation network M+P combined with multi-view geometry achieves higher accuracy.