Research On Mobile Robot SLAM Based On Deep Learning In Dynamic Environment

Simultaneous Localization and Mapping(SLAM)is an important technology for realizing computer vision positioning.It is widely used in the development of unmanned driving,humancomputer interaction and other fields,and has become a research hotspot.However,the current mainstream SLAM is mainly applied in a static environment,and a lot of work has been done on the optimization of accuracy;in a dynamic environment,the movement of objects in the scene will cause mismatches to the feature points during the pose estimation process.The robustness is not high enough and the operation efficiency is low,and a unified mature solution has not yet been formed to solve this problem.Therefore,this topic is aimed at researching the SLAM system in a dynamic environment,fusing the deep learning algorithm with the SLAM system,and improving the positioning accuracy of the optimized SLAM system without affecting the real-time performance of the system.First,it compares and analyzes the current research status of traditional SLAM and SLAM systems based on deep learning,and studies the impact of object movement in real environments on the pose estimation of visual robots and the reasons for the poor positioning accuracy of SLAM;and analyzes each structure of SLAM,Summarize the camera model,the principle and elimination method of the distortion caused by camera acquisition,compare the number of feature points and operating efficiency of five mainstream feature point extraction algorithms,and determine to apply ORB(Oriented FAST and Rotated BRIEF)as the feature extraction algorithm.In order to solve the problem of poor robustness of traditional SLAM in a dynamic environment,a Mask R-CNN convolutional neural network method based on deep learning is proposed to optimize the tracking of the visual front end.The ORB-SLAM and Mask R-CNN algorithms are effectively integrated to eliminate dynamic feature points,which strengthens the problem of poor positioning accuracy due to mismatching of dynamic feature points in the process of feature point matching;in addition,the residual Dynamic objects use multi-view geometric segmentation and epipolar geometric methods to remove the dynamic feature points of residual objects,which improves the accuracy of SLAM positioning.Finally,a comparative experiment was conducted with the ORB-SLAM2 algorithm in the public data set to verify the positioning accuracy and real-time performance of the optimized algorithm in this article;the Basler monocular camera was mounted on the departure 4 mobile robot,in the actual scene of people walking The camera positioning experiment was carried out,and the task of precise positioning of the camera was completed.The experimental results verify the effectiveness of the improved system based on the fusion of Mask R-CNN algorithm and ORB-SLAM in a dynamic environment.