Research On SLAM Technologies Of Mobile Robot Under Dynamic Environments

With the development of mobile robot technology,robots have gradually replaced human beings in the repetitive,heavy and dangerous working environments.SLAM technology is playing the key role for mobile robots to execute complex tasks including environment perception,path planning and human-computer interaction in unknown or complex environments.SLAM technology of mobile robot is mainly used in the field of handling robot,unmanned vehicle,UAV,feeding robot,navigation robot,floor sweeping robot and so on.In the indoor and outdoor environments of real applications,mobile robots are vulnerable to dynamic object interference,run in complex environment and so on,it is therefore necessary to research the localization and mapping of mobile robots in dynamic environment.This paper mainly focuses on the SLAM technologies based on vision and lidar in the complex dynamic environments,and aims to improve the localization accuracy and mapping quality of mobile robots when sufferring some interference factors,such as dynamic moving objects or changing background environment.On the basis of deep learning image semantic segmentation,a method of constructing a large-scale 3D semantic map is proposed.The main work and innovation are as follows:(1)A visual SLAM algorithm in dynamic environment is proposed.This algorithm uses the YOLOv3 model to detect the movable objects and filter the bad ORB feature points in the movable object area,and realizes an improved ORB-SLAM2 algorithm with certain real-time operation efficiency.Experiments were carried out on KITTI dataset and TUM dataset under the dynamic scene sequence,and the results showed that the localization accuracy of the algorithm in the outdoor environment is significantly improved compared with the original ORB-SLAM2,and the positioning accuracy in the indoor environment is close to Dyna Slam,but the speed is better than Dyna Slam.The algorithm achieves a speed of 14.3 FPS in the KITTI monocular dataset and 14.5 FPS in the TUM RGBD dataset.(2)A visual and lidar fusion algorithm for SLAM is designed and implemented.This algorithm is based on the synchronous acquisition of 2D images and 3D lidar point cloud data,and the high-resolution depth image is obtained by up-sampling and bilateral filtering,and the 3D coordinates of feature points are further calculated.Compared with the simple monocular/stereo visual SLAM algorithm,the localization accuracy of mobile robot can be further improved.Experiments applied in KITTI dataset and Apollo Scape dataset show that compared with the ORB-SLAM2 algorithm based on vision,the algorithm has significantly improved localization accuracy in most sequences of KITTI dataset and Apollo Scape dataset.(3)A novel 3D semantic map construction method is proposed and implemented.Our method fuses the recovered depth map with the pose information from SLAM localization.The 3D point cloud map reconstruction and semantic segmentation of robot motion scene were completed.The experimental results show that the 3D point cloud reconstruction is of good quality,dense and complete,and can recover scene information well,which is also better than the 3D reconstruction results of Dyn SLAM algorithm.Then,the Mask R-CNN algorithm was used to get the semantic segmentation of 2D images,and the semantic information was projected to the 3D point cloud map to achieve the 3D semantic point cloud map.We use octree structure to store and manage massive 3D semantic data,and provide environment semantic information for path planning and navigation of mobile robots.