We consider the special case of tracking objects in highly structured scenes. In the context of vehicle tracking in urban environments, we offer a fully automatic, end-to-end system that discovers and parametrizes the lanes along which vehicles drive, then uses just these pixels to simultaneously track dozens of objects. This system includes a novel active contour energy function used to parametrize the lanes of travel based only on the accumulation of spatiotemporal image derivatives, and a tracking algorithm that exploits longer temporal constraints made possible by our compact data representation; we believe both of these may be of independent interest. We offer quantitative results comparing tracking results to ground-truthed data, including thousands of vehicles from the NGSIM Peachtree data set.
See this youtube video for an overview of the system. You may also be interested in the poster [11MB; pdf]
This project is supported under NSF IIS 0546383: "CAREER: Passive Vision, What Can Be Learned by a Stationary Observer". Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.