Featureless Motion Vector-Based Simultaneous Localization, Planar Surface Extraction, and Moving Obstacle Tracking

Motion vectors (MVs) characterize the movement of pixel blocks in video streams and are readily available. MVs not only allow us to avoid expensive feature transform and correspondence computations but also provide the motion information for both the environment and moving obstacles. This enables us to develop a new framework that is capable of simultaneous localization, scene mapping, and moving obstacle tracking. This method first extracts planes from MVs and their corresponding pixel macro blocks (MBs) using properties of plane-induced homographies. We then classify MBs as stationary or moving using geometric constraints on MVs. Planes are labeled as part of the stationary scene or moving obstacles using MB voting. Therefore, we can establish planes as observations for extended Kalman filters (EKFs) for both the stationary scene and moving objects. We have implemented the proposed method. The results show that the proposed method can establish plane-based rectilinear scene structure and detect moving objects while achieving similar localization accuracy of 1-Point EKF. More specifically, the system detects moving obstacles at a true positive rate of 96.6% with a relative absolution trajectory error of no more than 2.53%.