Research On Semantic SLAM Combined With Deep Learning In Dynamic Environment

With the rapid development of artificial intelligence technology,Simultaneous Localization and Mapping(SLAM)has become a key technology in intelligent autonomous robots,and it has also been widely used in autonomous driving and augmented reality.Static environment is a prerequisite for most of visual SLAM,which limits the practical application of most existing SLAM systems.It can't work well when moving objects enter the camera's field of view.At the same time,the random movement of the moving object in the environment will contaminate the constructed global point cloud and make it lose its visual significance.While traditional SLAM algorithms are mostly completed using geometric information,intelligent machines need to interact from the semantic level.Therefore,the SLAM system with semantic information has become a new breakthrough with the rapid development of deep learning.For these problems,this paper studies the semantic SLAM combined with deep learning in a dynamic environment.The specific research results are as follows:(1)Aiming at the problem that moving objects will interfere with the pose calculation,a double-constrained dynamic matching point detection algorithm combined with semantic information is proposed.A deep network BlitzNet capable of simultaneous target detection and semantic segmentation is introduced to adapt to SLAM.The feature point extraction is improved by a multi-layer multi-threshold corner extraction algorithm,and the quadtree is used to optimize the management of feature points to maintain a certain scale.Depth constraints and epipolar constraints of region division are combined to classify dynamic and static matching points in feature point matching.On this basis,the dynamic target detection can be completed.Experiments show that the algorithm can effectively detect dynamic matching points and dynamic targets for indoor dynamic scenes.(2)Aiming at the problem that moving objects will cause pollution to point cloud,a registration scheme of point cloud based on classified denoising and weighted ICP is present.Residual isolated points in point cloud are removed using sliding windows and histogram statistics,and residual noise blocks of the interactive target are processed with semantic and morphological.The ICP algorithm was improved,and the point cloud control point sampling was completed by triple-condition screening,which reduced the calculation cost.Corresponding weights are added in the ICP solution process to complete point cloud registration.Compared with two traditional registration algorithm,the weighted ICP has good registration effect and calculation efficiency.The proposed point cloud stitching algorithm is used to complete the point cloud stitching in a dynamic environment,which proves the feasibility of the algorithm.(3)A semantic ORB-SLAM2 algorithm adapted to the dynamic environment is proposed.The structure and the shortcomings of the ORB-SLAM2 algorithm are analyzed.Deep network BlitzNet is integrated into the system,combined with a dual-constrained dynamic point detection algorithm and a mapping scheme in a dynamic environment,to improve the ORB-SLAM2 system.The improved ORB-SLAM2 algorithm is evaluated from the two aspects of localization and mapping.Experiments shows that improved ORB-SLAM2 algorithm can improve the positioning precision in the dynamic environment and eliminate the smear generated by the dynamic objects in the point cloud.The reconstructed clean point cloud contains simple semantic information,which has visual significance.