Dense 3D SLAM in Dynamic Scenes Using Kinect

In this paper, we present a dense 3D SLAM method for dynamic scenes consisting on building, in real-time, a 3D map of the scene using Kinect. The method starts by segmenting and removing moving objects from the scene in order to avoid mismatches in the alignment step, then, calculates the scene current camera pose for each new acquisition. This method has the advantage of producing a dense map through the use of all pixels of the RGBD camera in order to achieve higher pose accuracy. The method also includes a loop closure detection thread to detect and merge duplicate regions. Quantitative evaluations using a various sets of scenes and benchmark datasets show that the proposed method produces a real-time 3D reconstruction with higher accuracy and lower trajectory error compared to the state-of-the-art methods.