Dense Simultaneous Localization And Mapping Based On Scene Segmentation

Compared with the traditional Simultaneous Localization and Mapping,SLAM based on depth camera can directly capture the depth value of image points,thus can carry out real-time dense mapping,which has attracted wide attention in recent years.However,when there are dynamic objects in the scene,the moving objects in the scene will affect the accuracy of localization and the global consistency of mapping.Robust SLAM in dynamic scenes is a challenging task.Aiming at the problem of dynamic scene,on the idea of detecting and eliminating dynamic targets,a dense SLAM method based on scene segmentation is proposed.It can construct a dense scene with global consistency while detecting moving objects.In the frontend camera localization part,firstly,the scene data collected by the depth camera are preprocessed,and then clustered using K-means segmentation algorithm.After initializing the probability that each cluster block belongs to the dynamic target,the depth and gray information are used to establish the solution model of the camera pose and the dynamic probability of the cluster block,finally,using the iterative re-weighted least squares algorithm to solve the problem.In the back-end mapping part,the stability parameters are introduced into the map model by using the dynamic probability results of clustering blocks in the localization part,and the basic surface representation method is improved.A dense mapping method combining dynamic probability is proposed to construct a dense scene containing only static targets.The experimental results of high dynamic scenes in our lab environment show that the proposed method can effectively segment the dynamic objects and build a globally consistent dense map.The experimental results in different scenes of public datasets show that,for the accuracy of camera location,our method achieve the same results as the stateof-art methods in the static and low dynamic scenes and improves the accuracy in the dynamic scenes.