Research On Mobile Robot SLAM Algorithm Based On Image Semantic Segmentation

With the wide application of mobile robots,SLAM(Simultaneous Localization and Mapping),as the core algorithm of mobile robots autonomous navigation and environment interaction,which is gradually attracting the attention of academic researchers and engineers.The current SLAM algorithms are generally based on the geometric characteristics of the environment,their accuracy has reached the bottleneck,and its robustness needs to be improved.With the emergence of new application scenarios such as unmanned vehicle and human-machine collaboration,robots are required to have the ability to perceive environmental information while having semantic information perception capabilities.With the application and breakthrough of deep learning technology in computer vision,it is a research hotspot in the current academic circles to assign this ability to robots.In order to solve the problem of robots acquiring semantic information in real time,a real-time semantic segmentation algorithm combining spatial information and semantic information is proposed.From the understanding of semantic segmentation task,semantic segmentation is decomposed into a task combination of spatial information localization and semantic information classification.The spatial information branch,the semantic information branch and the information fusion module are proposed,and the effectiveness of each module is verified by feature visualization.The algorithm only needs 16 ms to infer a picture with a resolution of 640×480 on GTX1080 Ti,and achieves 71.3% mIoU results on Pascal VOC dataset,which realizes the robot cognitive environment.In order to improve the robustness of the traditional SLAM algorithm,a SLAM algorithm combining semantic information is proposed.By adding semantic information,we solved the problem that the traditional RANSAC(Random Sample Consensus)algorithm does not consider the intrinsic link of the sample and the assumption of the sample relationship is too strong and does not conform to the law of rigid body motion.Using the rigid body motion law as a priori knowledge,the initial pose estimation robustness of the robot is improved.The feature point filtering method is used to decouple the various motion states in the multi-object motion scene,and based on the feature point semantic information association,the data association consistency of the SLAM algorithm in the tracking and loopback detection process is enhanced,and the robot's pose estimation accuracy.Finally,the experimental platform of mobile robot is built,and the standard data set and real scene experiment verification are carried out on the proposed algorithm.The experimental results show that the proposed algorithm has more significant improvements in the root mean square error of the absolute estimation and relative estimation than the current ORB-SLAM2 algorithm in many scenarios,and has good generalization performance.