Award

The iRobot team proposes to develop a deployable pedestrian detection module that will analyze a monocular video stream and identify all instances of visible pedestrians. To address the problem of pedestrian detection, we will employ data-driven approaches that help recognize specific appearance and motion patterns generated by moving humans. We will also develop detection algorithms that can leverage the decentralized use of multiple mobile sensor platforms and improve detection accuracy. To reduce false positives while also maintaining an acceptably high detection rate, we will explore mechanisms for exploiting scene context. We will also explore methods to obtain useful context. Our primary technical objectives in Phase II include developing a pedestrian detection algorithm that performs competitively with other published pedestrian detection algorithms; exploring mechanisms for integrating vehicle telemetry information and other contextual information with the detection results; evaluating the ability of our integrated detection solution to approach the detection performance of existing compute-intensive systems for pedestrian detection; operating the pedestrian detection algorithm onboard a tactical UGV platform without significant sensor or hardware modifications; and evaluating the pedestrian detection algorithm performance onboard a tactical UGV platform under urban outdoor conditions involving groups of five or more targets.