Detecting and Tracking Multiple Moving Objects from a Moving Platform

We propose to demonstrate a"Tracking with Planning"module for tracking objects in a highly cluttered environment from a moving platform. This system is designed to track objects with multiple long occlusions, by anticipating where and how the objects might move. The main innovations are 1) actively anticipating objects'goals and intentions, 2) using motion planning to create multiple plausible predictions on objects'trajectories, and 3) online learning of visual features for discriminating similar nearby objects. In phase 1 we have demonstrated that the"Tracking with Planning"approach can overcome many problems associated with traditional tracking algorithms, such as drifting into background after multiple occlusions. We improved the tracking precision from 56% using Kalman filter to 87% using"Tracking with Planning."In phase 2 we propose to develop an active goal prediction system with multiple path generation. We plan to integrate bottom-up information from tracking and contextual visual information with top-down planning-based reasoning to create a robust tracking system. We can adapt our modular design to different operating environments from outdoor urban environments to indoor office settings. We plan to evaluate this system on a mobile ground platform as well as a simulated unmanned helicopter in preparation for flights on a tactical autonomous helicopter.