Dense Map Construction And Updating For Dynamic Environment

Simultaneous Localization and Mapping(SLAM)stands a central part of autonomous navigation.In the framework of SLAM system,building dense maps,the core parts of SLAM,provides more complete environmental information,which benefits to localization,navigation and obstacle avoidance.Most of the current dense mapping is based on static environment.However,in practice,the presence of dynamic objects such as pedestrians and vehicles results in poor qualities of dense map building.Therefore,it is important to solve the problem of dense mapping in dynamic environment.Based on the previous work,this paper designs a low-cost and efficient detection method of dynamic objects which relies on no prior.Combined with the robust dynamic dense visual odometry,it finally realizes the building and updating of dense map.Our work is as follows.(1)A novel dynamic object detection algorithm independent of deep learning is proposed.Traditional dynamic environment mapping mostly relies on deep learning to realize pedestrian detection and elimination.However,deep learning requires high hardware cost and poor real-time performance.In this paper,we design a fast segmentation method of front and back scenes,without a priori information,using RGBD video stream to achieve real-time and accurate detection and segmentation of dynamic objects.Experiments show that the algorithm can detect dynamic objects effectively and reach a good robustness.(2)Combining with the dynamic object detection module,a robust dynamic dense visual odometry is proposed.In this paper,a multi-level residual model was constructed by using a binary matrix which records dynamic and static information.The residual constructed by dynamic pixels in each residual model layer were eliminated,and a more accurate and robust camera pose was obtained by least square optimization.Experimental results show that the proposed dynamic dense vision odometry has higher accuracy than other algorithms in estimating the camera pose.(3)Considering the complexity of the motion state of dynamic objects in the scene,method of map fusing and updating was proposed to build a static background map.In this paper,a depth-based background model is proposed to identify the wrong elements of the map model,which helps to update the map dynamically and improve the accuracy of dense map building.The experimental results show that the map model constructed by the algorithm proposed in this paper has low error.