Research On Indoor Semantic Static Map Construction Based On Multi-scale Information Fusion For Visual SLA

The emergence of many intelligent devices such as robots and intelligent industrial vehicles has liberated productivity and changed the way of life.Simultaneous Localization and Mapping(SLAM)systems play an indispensable role,and are often used to estimate pose and build environmental consistency maps,which are widely used in home and factory production environments.However,with the emergence of new production and living needs,smart devices are required to have the ability to perceive and understand the surrounding environment,and assist smart devices to complete behavior-level tasks such as navigation and obstacle avoidance,human-computer interaction,and object manipulation.Therefore,building a environmental map with semantic information,perceptible and comprehensible information becomes very important for smart devices.This paper takes the vision system of home care robot as the research background,the aims are: in the indoor dynamic scene,through dynamic object culling and image multi-scale information fusion,make certain innovations and expansion to the existing ORB-SLAM3 framework to further construct static semantic maps and octree maps that can be used for upper-level interactive applications and lower-level motion control of robots.The main research contents and work results are as follows:(1)For visual SLAM in dynamic scenes,due to object motion,it faces a problem that the feature point data is incorrectly associated,and the pose cannot be accurately estimated.This paper proposes an algorithm that uses the semantic information obtained by the instance segmentation network and the global dense optical flow to form mutual constraints,and generates a dynamic and static mask to eliminate dynamic feature points,which improves the robustness and tracking accuracy of ORB-SLAM3 in dynamic scenes.(2)In the way of image fusion,the information of different scales of the image is fused to explore and improve the representation effect of the environment map.And an algorithm for multi-scale fusion of images using wavelet multi-level transform is proposed: in the sub-band fusion process,the low-frequency sub-band fusion directly uses the low-frequency contour obtained by the first-level decomposition to approximate the sub-band;the high-frequency sub-band fusion rules are compared to select gradient feature-based methods,and then the multi-scale residual pyramid constructed by each sub-band is involved in the final fusion process to form an effective image multi-scale information fusion scheme after blurred.The algorithm aims to improve the quality of map representation while giving the map perceptible and understandable information.(3)On the basis of dynamic object culling and image multi-scale information fusion proposed,the two-dimensional semantic information after fusion of image multiscale information and the key frame pose information generated by ORB-SLAM3 are combined to construct indoor static semantic map and octree semantic map,which expands the semantic dense mapping function of ORB-SLAM3.The mapping validity and feasibility of the total system constructed are confirmed by the dataset experimental comparison verification and the actual environment operation test,and it will provide assistance for the upper-layer interactive application and lower-layer motion control of the home care robot.