Research On Dynamic Environment Visual SLAM Using Deep Learning

The gradual application of Synchronous Localization and Mapping(SLAM)technology in the process of robot intelligence reflects the further improvement of technological development level.This technology includes basic positioning and mapping functions,providing support for indoor navigation and path planning of robots.When using a camera as its underlying hardware,the research field formed is called visual SLAM.Currently,most visual SLAM systems can operate stably based on the assumption that the operating environment is static or rigid.However,the real environment is complex and dynamic,and dynamic objects in the scene can bring significant errors to the system’s pose estimation,reducing the system’s positioning performance,resulting in poor visibility of the generated point cloud map.And the lack of necessary semantic information further affects the effect of map construction.In response to the above issues,this thesis is based on the ORB-SLAM2 system and applies a deep learning based object detection model to the visual SLAM system,conducting research on visual SLAM in dynamic environments.The Ghost-YOLOv5s and the improved polar geometry constraint algorithm are combined to filter the dynamic feature points,improve the positioning accuracy of the system,and construct a 3D semantic point cloud map without dynamic object information in dynamic scenes.The main content includes:(1)When using object detection for dynamic feature point filtering,the existing deep learning based object detection model YOLOv5s convolution layer has a large number of parameters and high computational requirements,which is not conducive to deployment in mobile or embedded edge environments when combined with SLAM systems.Therefore,the Ghost-YOLOv5s object detection model is proposed to be applied to SLAM systems for dynamic feature point filtering,replacing most of YOLOv5s convolutions with Ghost convolutions,Make the model lighter and more real-time.(2)In response to the problem of excessive or insufficient filtering caused by the unreasonable selection of the distance threshold range from the matching point to the polar line in traditional polar geometry constraint algorithms for secondary filtering of dynamic feature points,an improved polar geometry constraint algorithm is proposed to filter out secondary dynamic points.By combining the feature points obtained from static object boxes through object detection,the threshold solution method is changed to accurately distinguish between dynamic and static points in the scene,Filter out real dynamic points and improve the excessive or insufficient filtering of feature points.(3)Aiming at the problems that the visual ORB-SLAM2 system in the dynamic environment uses its original key frames to build a map that has a large number of dynamic object ghosts and redundant information,which leads to poor readability of the map and lack of necessary semantic information,a 3D semantic map building algorithm in the dynamic environment is proposed.Firstly,the key frames are jointly filtered by interval filtering,image similarity detection and depth data detection.Then,the semantic information obtained by the object detection model Ghost-YOLOv5s in this thesis is applied to filter the dynamic information in the scene,and the dense point cloud and octree map in the dynamic scene are constructed.Finally,the semantic information and point cloud are fused to construct a 3D semantic point cloud map.(4)Build relevant experimental platforms and conduct experimental analysis on the algorithm proposed in this thesis.Dynamic object detection experiments have shown that Ghost-YOLOv5s requires fewer model parameters and computational power compared to YOLOv5s,which is beneficial for deployment in mobile or embedded edge environments.The dynamic point filtering experiment of SLAM system shows that Ghost-YOLOv5s+improved polar geometry constraint algorithm has a good effect on filtering dynamic points,which can significantly improve the filtering of feature points over or under,and has stronger real-time performance compared to other classic dynamic visual SLAM systems;The 3D semantic map construction experiment shows that the constructed semantic map has no interference from dynamic object ghosting,less redundant information,strong readability,and obvious semantic features.