Realtime Visual and Point Cloud SLAM

The availability of affordable RGB-D cameras like Microsoft Kinect can improve VSLAM applications, object 3D modeling and reconstruction of indoor environments, through the use of dense, synchronized depth and color images. The high frame rate of such devices isn’t exploited so far, since they require both fast and accurate algorithms for real-time registration. In this paper we present a technique that performs generalized ICP on two frames in typical times of 10 ms. The technique uses an efficient bundle-adjustment framework, so that we can combine ICP with visual feature matches, both for frame-frame matching, and overall global adjustment.