Detecting and Tracking Multiple Moving Objects from a Moving Platform

Successful detection and tracking of multiple moving objects from a ground robot will greatly facilitate important military tactical missions such as path planning, while operating safely amongst humans. Despite extensive research over past decades, robust and efficient visual detection and tracking from a moving platform in unconstrained real-world environments remains challenging. Most existing algorithms make specific assumptions. SET Corporation proposes to develop a persistent multiple moving object detection and tracking system for an autonomous vehicle equipped with stereo cameras with optional LIDAR input. Our system integrates motion & stereo computation, object recognition, tracking, and trajectory analysis to efficiently and reliably detect multiple discrete moving or static objects for a long period of time. The computation is based on a unique competition and collaboration framework. At local level, 3D spatial objects occupying the same pixels in 2D image planes compete with each other to best explain both spatial and temporal evidence. At global long term trajectory level, object trajectories that are easy to track and maintain high confidence are used to guide the association of trajectories with less confidence. The collaboration is formulated using motion and appearance context. The resulting approach can initialize automatically and track a large and varying number of human and vehicular objects over long periods and through complex scenes with clutter, occlusions, and large-scale background changes. Moreover, an in-situ tracking evaluation scheme is utilized to systematically assess the tracking performance.