Research On Visual Perception Algorithm Of Mobile Robot Based On Semantic ORB-SLAM2

The prerequisite for a mobile robot to successfully perform a target navigation task in an unknown environment is to accurately perceive and understand the surrounding environment.The robot's autonomous localization and mapping algorithm in an unknown dynamic environment is improved,and useful information in the environment from various aspects is mined.By detecting dynamic objects and eliminating the moving point noise,a high-precision trajectory is generated.Through the key frame projection,the static background of the current image is restored,so that the mobile robot has the ability of "perspective" and can perceive the scene information behind the occluded part.The 3D point cloud semantic map is constructed by fusing the 2D image target detection results,so that the robot can obtain the object category and location information in the scene according to the built map.Through the point cloud segmentation,the probability occupation map is established dynamically,which can provide reference information for mobile robot to avoid obstacles and plan routes after identifying the target object,and further improve the robot's ability to perceive the environment.The research goal of this paper is to solve the problem of autonomous positioning and perception of environmental information for mobile robots in unknown environments.The research idea is to combine visual simultaneous localization and mapping(SLAM)with deep learning to realize the visual positioning and environmental perception functions of mobile robots based on semantic O RB-SLAM2.The work done is as follows:(1)Research and improve the ORB-SLAM2 algorithm.Before the feature point tracking and matching,the priori dynamic object in the image is segmented by using the deep convolution neural network Mask R-CNN,and other non priori dynamic regions associated with the priori dynamic object are also segmented by the method of multi-view geometry.Therefore,it can eliminate the noise of moving points in the environment,reduce the error caused by the noise of moving points in the calculation of camera pose,and make the map constructed more accurate.(2)The image occlusion repair algorithm is researched to restore the background area occluded by dynamic objects on the image to realize the restoration of the dynamic environment to the static environment.By traversing the key frame database,the pixels on the static background area of the key frame are projected over,and the background occluded by dynamic objects in the current frame image is repaired.(3)The target detection algorithm SSD based on deep learning is studied.The semantic information of 2D image is obtained by Mobilenet?v2-ssd-lite network model,and the generated 3D map points are semantically segmented to achieve the fusion of 3D object geometric information and semantic information,so as to build a 3D point cloud semantic map.(4)Investigate the OctoMap principle of the probabilistic occupation map of the three-dimensional environment.In the ORB-SLAM2 mapping thread,dynamically generate the probabilistic occupation map OctoMap so that the mobile robot can determine whether a certain place can pass safely based on the occupation probability map.(5)Experiments are performed on the algorithm designed in this paper.The TUM RGBD dataset is used to check the implementation effect from three aspects: algorithm function,algorithm accuracy,and algorithm time to verify the feasibility and accuracy of the algorithm.The experimental results prove that the work done in this paper improves the function of ORB-SLAM2 and improves the accuracy of the algorithm.In terms of time,different algorithm strategies should be adopted according to different environments.Finally,according to the deficiencies of the algorithm found in the experiments,future research work that can be further improved is prospected.