6D SLAM with RGB-D Data from Kinect
We quickly acquire colored 3D models of objects and indoor scenes with a hand-held Kinect camera. We extract SURF features from the camera image and localize them in 3D space using the depth image. We match these features between every pair of acquired images, and use RANSAC to robustly estimate the 3D transformation between them. To achieve real-time processing, we match the current image only versus a subset of the previous images with decreasing frequency. Subsequently, we construct a graph whose nodes correspond to camera views and whose edges correspond to the estimated 3D transformations. The graph is optimized online to reduce the accumulated pose errors. This allows us to deal with loop closures, i.e., after the user has moved the camera around an object and returned to a known pose.