Dense SLAM Based Tracking And Reconstruction For Multi-Objects

The technology of dense SLAM(Simultaneous Localization and Mapping)based tracking and reconstruction for multi-objects has important research significance and application value in the fields of robot navigation,human-computer interaction,augmented/virtual reality and so on.However,the system often cannot detect dynamic objects correctly because of the objects being occluded or leaving the camera field of view,and the tracking process will fail.In order to achieve robust long-term tracking of scene objects,this paper proposes an unsupervised long-term tracking and 3D reconstruction method for dynamic objects,which can distinguish the static background and each dynamic foreground object independently.The main work and contributions are as follows:1.An accurate target tracking strategy combining two-dimensional vision and depth information is proposed,which can detect and track the dynamic scene objects correctly in the case of large-area occlusion or long-term out-of-view.First,the motion pose of each data point is calculated by a motion pose estimation energy function,then the motion pose is combined with the input RGB-D information to perform the edge optimization based segmentation,and then the data points with similar properties are gathered into the same set,and an independent tracking area is generated for each dynamic object.The experimental results show that the segmentation region generated by this method fits the boundary of the object,and can always keep a unique label for each object to achieve robust tracking.2.A 3D model reconstruction method based on adaptive fusion and update strategy is proposed,which provides a positive feedback for the calculation of motion pose.This method needs the motion pose of the previous frame as a piece of prior information,and then combines the pose data with the input information to generate the 3D reconstruction model,and finally combined with the adaptive update strategy to optimize the surface of the reconstruction model.The experimental results show that this method can obtain accurate motion pose results,and can generate the dense reconstruction model that close to the real surface.In the end,this paper summarizes the work and puts forward the further work goals,including reducing the memory consumption and improving the system operation efficiency,and increasing the type of input devices to expand the application scenarios.