Research On RGB-D SLAM System In Dynamic Scene

With the development of big data and artificial intelligence technology in recent years,intelligent robots can perceive the surroundings independently like human beings,and they can complete more and more complex intelligent tasks.The maturity of SLAM technology enables the machine to determine the specific location itself when it works,and build a map of the environment it passes through at the same time.Most of the existing visual SLAM algorithms are proposed under the assumption of static environment,but such ideal environment is difficult to be guaranteed in reality.During the operation of the robot,there might be moving objects appearing in its filed of view,and these dynamic objects change the layout of all objects in the environment by contacting others and moving them.In this case,not only would such visual SLAM algorithms inaccurately estimate the pose of the camera,the environmental maps they build would contain serious dynamic noise,resulting in abnormal operations of the robot in the dynamic scene.In order to solve the problem,this paper studies the RGB-D SLAM system in dynamic scene,and it mainly completes the following research work:(1)Aiming at the interference of moving objects in image features in dynamic environment,a high-practicable feature detection algorithm based on joint constraint is proposed.Firstly,the pre-trained YOLACT instance segmentation network is used to segment the color image taken by the RGB-D camera.The image is divided into potentially dynamic region and environmental region according to prior knowledge,so that the features in potentially dynamic region will be tested for depth consistency and distance consistency.The features which do not meet the requirements would be removed from the previously extracted results.Then,the quadtree data structure is utilized to extract features in the area outside the dynamic object bounding boxes,so as to obtain image features which are reliable,sufficient and uniformly distributed.(2)In order to solve the problem that a lot of noise exists in the map constructed by visual SLAM algorithm in dynamic scene,a map denoising algorithm based on dynamic complement is proposed.The potential dynamic mask with depth consistency is generated by depth search in the adjacent area of the moving object.The semantic mask of the moving object is developed by region segmentation and semantic fusion for the potential dynamic mask,so that the dynamic instance mask is obtained with more perfect coverage for the real object.At the same time,motion re-recognition of all objects in the environment is carried out by combing epipolar constraint with depth matching,so the new moving objects in the image could be located accurately.Finally,the dynamic noise in the point cloud map and octree map can be effectively filtered after dynamic complement,thus a more accurate and low-noisy environmental map can be built.(3)This paper integrates dynamic processing module and global mapping module on the basis of ORB-SLAM2 algorithm,which constructs a dynamic semantic RGB-D SLAM system.The dynamic processing module uses the instance segmentation network and depth measurement information to effectively eliminate the influence of moving objects in the dynamic scene,and it improves the accuracy of pose estimation in the tracking module by extracting the high-practicable features in the image.Meanwhile,static and stable landmark map points can be generated in local sparse map.Utilizing dynamic instance mask complement and object motion re-recognition in the global mapping,the dynamic noise in the global environmental map is effectively filtered and the interference of dynamic noise in the octree voxel occupation state is avoided.