Research On SLAM Mapping For Indoor Working Robots In Dynamic Scenarios

Simultaneous Localization and Mapping technology is the key to realize environment perception and autonomous motion of mobile robots.With the increasing maturity of SLAM technology,most open SLAM algorithms have basically realized autonomous localization and mapping in static environment.At present,the visual SLAM system with RGB-D camera as sensor has made great progress,but there are still some problems,such as low tracking efficiency of key frames and low loopback detection efficiency of single visual odometer in different scenes.Moreover,when there is a moving object in the scene,the uncertainty existing in the real scene will interfere with the feature matching of SLAM system,resulting in the accuracy of SLAM system positioning and mapping is reduced,so that its practical application is limited.Therefore,in view of the above problems,the visual SLAM system of the operating robot in the indoor dynamic scene is studied in this paper,and map reconstruction is carried out.The main research contents of this paper are as follows.(1)In view of the low adaptability of a single visual odometer in changeable scenes,this paper optimized the visual odometer and proposed an improved optical flow tracking algorithm.According to the different motion states of the camera in the scene,the visual odometer suitable for the current scene was selected to improve the adaptability of the algorithm to different environments.After verification on TUM data set,it can be concluded that the proposed algorithm can better adapt to different scenes than the single visual mileage calculation method.(2)Aiming at dynamic scenes,this paper studies an image moving region extraction method based on semantic segmentation network,which removes dynamic objects in the environment by segmenting potential moving objects in the scene.The3 D semantic scene map is reconstructed in the static scene after removing the dynamic object,and the global point cloud map and 2D raster map are constructed respectively in the data set and the real scene,which verifies the construction efficiency of the environment map of the improved SLAM system in this paper and the robustness of the map construction in the dynamic environment.(3)Aiming at the problem that the traditional visual word bag model is easy to be limited by the size of the visual word bag,which leads to the low accuracy of loopback detection,this paper improves the selection of loopback frame in the loopback detection module based on the deep learning algorithm and optimizes the problem of point cloud error matching in the back-end optimization.First,in the loopback detection part,the selection strategy of loopback frame is improved,the key frame sequence is established for the key frame,and only the dictionary of the key frame in the sequence is matched,so as to accelerate the feature search speed.Then,after judging the similarity and continuity of key frames in the sequence,a more reliable loop frame can be obtained.In the part of back-end optimization,an idea of improving NICP matching algorithm is proposed to solve the problem of point cloud mismatching caused by triangulation.The loopback frame optimization experiments were carried out on TUM data set and the whole SLAM system after improvement.The experimental results show that the improved loopback frame selection algorithm can effectively improve the efficiency and accuracy of loopback detection.Finally,from the perspective of the overall SLAM system,the improved SLAM system in this paper has achieved high accuracy in positioning and mapping.