Event Location: NSH 1507
Bio: Robert T. Collins received the Ph.D. degree in Computer Science from the University of Massachusetts at Amherst in 1993. He is an associate professor in the Computer Science and Engineering Department at The Pennsylvania State University, where he co-directs the Lab for Perception, Action and Cognition (LPAC). Prior to joining Penn State in 2005 he was on the research faculty of the Robotics Institute at Carnegie Mellon University, where he worked for over a decade on DARPA projects such as VSAM, HID and VIVID, and helped to develop the CBS EyeVision System, for which he holds three joint patents with co-inventors at CMU. His current research interests include video scene understanding, automated surveillance, human activity modeling, and multi-target tracking. He is a senior member of the IEEE, a member of the IEEE Computer Society, a member of the Computer Vision Foundation, and an associate editor for the International Journal of Computer Vision.

Abstract: n the first stage of the “detect-then-track” paradigm, an object detector is run on each frame of video to hypothesize objects of interest. This is followed by a second, data association stage, to link the detections into multi-frame trajectories. This second stage of multi-frame data association is of particular interest, as it is a combinatorial optimization problem of significant complexity. We argue that recent vision-based approaches rely too heavily on object appearance cues to solve this problem, to the point of ignoring motion characteristics. One example is the recent network flow formulation where the number and best set of trajectories can be solved optimally using min-cost flow in polynomial time. Although an exciting result, the approach relies on being able to factor trajectory cost functions into the product/sum of pairwise costs on each frame-to-frame link. This limits evaluation of geometric track quality to terms based only on distance traveled between frames, e.g. shortest paths, and does not allow for higher-order smoothness constraints that are functions over three or more frames, such as piecewise constant velocity. The lack of regularizing motion models has a detrimental effect on quality of the trajectories found when appearance information is weak (e.g. thermal) or nonexistent (e.g. radar blips), and/or when the detection frame-rate is low. This talk will present recent work that combines elements from network flow into the more traditional, but NP-hard, multidimensional assignment formulation, resulting in efficient algorithms capable of finding high quality approximate solutions to the multi-frame data association problem under higher-order cost functions.