Visual Odometry and Mapping for Indoor Environments Using RGB-D Cameras

RGB-D cameras (e.g. Microsoft Kinect) offer several sensing capabilities that can be suitable for Computer Vision and Robotics. Low cost, ease of deployment and video rate appearance and depth streams are examples of the most appealing features found on this class of devices. One major application that directly benefits from these sensors is Visual Odometry, a class of algorithms responsible to estimate the position and orientation of a moving agent at the same time that a map representation of the sensed environment is built. Aiming to compute 6DOF camera poses for robots in a fast and efficient way, a Visual Odometry system for RGB-D sensors is designed and proposed that allows real-time position estimation despite the fact that no specialized hardware such as modern GPUs is employed. Through a set of experiments carried out on publicly available benchmark and datasets, we show that the proposed system achieves localization accuracy and computational performance superior to the state-of-the-art RGB-D SLAM algorithm. Results are presented for a thorough evaluation of the algorithm, which involves processing over 6, 5 GB of data corresponding to more than 9000 RGB-D frames.