Toward 1 Million Miles of Pedestrian Detection and Avoidance: Truly Robust Multisensor Safety Systems
Assistant Professor of Mechanical Engineering, College of Engineering
Associate Professor of Naval Architecture and Marine Engineering and Associate Professor of Electrical Engineering and Computer Science, College of Engineering
Research Process Manager
The project constructed a multisensor fusion-based approach for pedestrian and cyclist detection and a real-time optimization-based approach for collision avoidance. It used multiple sensors to construct an algorithm to detect and predict the motion of pedestrians and cyclists that is robust to occlusions from any single sensing modality. This algorithm was trained upon a realistic dataset drawn from existing Mobility Transformation Center (MTC) testbeds that is orders of magnitude larger than any existing dataset. The project also constructed a real-time optimization scheme to generate controls that avoid collisions with the predicted locations generated by the first algorithm to illustrate the utility of the predictions.
(1) A 10,000-hour annotated dataset of pedestrian and cyclist motion using the Safety Pilot Model Deployment dataset, which includes data from inclement weather; (2) a multisensor fusion-based approach to detect and predict the motion of pedestrians and cyclists; and (3) a real-time numerical optimization scheme that synthesizes controller interventions, which can safely avoid probabilistic predictions of pedestrian motion.