Real-time dense scene flow estimation using a RGB-D camera

In this paper, we present a novel framework for dense scene flow estimation using range data from a RGB-D camera. The Lucas/Kanade optical flow technique is extended to three dimensions for estimating dense scene flow. All of the computation is achieved in real time on an AscTec Pelican quadrotor onboard processor. One of the main ideas for our algorithm is to detect and predict the velocity of moving objects from the camera view. To achieve sufficient efficiency for real-time applications, we take advantage of the integral image technique to compute the value of arbitrary rectangular windows quickly. Experimental results of dense scene flow are shown in all 3 axes. Quantitative results are shown and analysed with different resolutions and various lighting conditions.