Robust Motion Estimation with RBG-D Cameras

Estimating vehicle motion using vision sensors in real time has been greatly explored in the past few years due to speed improvements and advances in computer hardware. Six degree of freedom motion estimation using vision information is desirable due to a vision sensors low cost, low power requirements and light weight and for the quality of the solutions that can be obtained using few assumptions about the environment. However, cameras have the downside of not providing good estimates when visual features are sparse or not available. Also, there are problems with changes in lighting and when light is low or unavailable. Laser scanners have been shown to be robust in these situations. We view an RGB-D sensor as providing three complimentary modalities that are useful for providing motion estimation solutions: a monocular camera, a 3D point cloud and the combination providing RGB-D information. Obviously motion estimates produced using the combined sensor information are best. However, there are times when information from both sensors is not available. The monocular camera remains useful when depth information is absent or insufficient, like in a large room, down a long hallway or outdoors. The 3D point cloud may still be available when there is insufficient light to utilize the RGB image. The approach described in this work seeks to take advantage of all three of these sensor modalities to provide a more robust motion estimation solution.