Research On Mapping Technology Of Indoor Mobile Robot Based On Dynamic Target Detection

Simultaneous Localization and Mapping(SLAM)is a key technology for mobile robot perception.The robot senses the surrounding environment by carrying a camera or lidar,locates itself and constructs an environment map for navigation and path planning.This paper focuses on the mobile robot using RGB-D camera as the sensing system in indoor dynamic scenes,and solves problems such as low positioning accuracy,weak tracking stability of feature points and dense mapping i n the dynamic scenes of visual SLAM algorithm.Firstly,in order to determine the position of dynamic objects in the imag e,a dynamic object detection algorithm based on polar constraint and semantic segmentation is designed in this paper.Dynamic object detection algorithm is mainly composed of two modules: dynamic feature point screening and semantic segmentation.In order to improve the efficiency of the algorithm,dual-thread parallel program design is adopted,and image data is input into two modules for processing.The dynamic feature point screening module establishes data association of adjacent image frames through LK optical flow tracking,and uses polar constraint to screen dynamic feature points.The semantic segmentation module designs a lightweight semantic segmentation network,which determines dynamic objects according to the ratio of dynamic feature points.Secondly,in order to improve the stability of feature point tracking process and reduce the time consuming of feature point extraction and descriptor calculation,a fast visual odometer based on feature optimization is designed in this paper.The main input data of visual SLAM system is feature points.In order to improve the quality of feature points and avoid using low-quality feature points in weak texture region,this paper designs a feature optimization algorithm to extract high-quality feature points.In order to reduce time consumption,this paper uses CUDA parallel computing architecture to design an accelerated algorithm for feature point extraction and descriptor computation.After matching the feature points,the pose was solved,and the reprojection error was nonlinear optimized to obtain the high precision pose.Thirdly,in order to obtain high quality environment map,this paper designs static environment map building algorithm based on local map Mosaic.Considering that the depth images collected by RGB-D cameras may have holes under the influence of the environment,this paper designed a depth image restoration and noise reduction algorithm based on the numerical distribution of pixel s in the field of holes.In the aspect of environment map construction,dynamic objects are filtered from the key frame image,the point cloud map of single frame image is constructed using the key frame and point cloud filtering is carried out.The point cloud map is transformed to the world coordinate system accord ing to the position and pose relation,and the point cloud map is transformed to the octree map which can be used for navigation.Finally,in order to verify the effectiveness of the proposed algorithm accurately and objectively,the proposed algorithm is evaluated respectively in the standard data set and the actual environment.In the evaluation process,static scenes,low dynamic scenes and high dynamic scenes from the TUM dataset are selected for testing,and experimental links of linear motion and clos ed-loop motion are set in the actual environment.Through comparative experiments,the effectiveness of the proposed algorithm in dynamic scene localization and mapping is fully verified.