Multi-Camera Tracking
People
Scott Spurlock Richard SouvenirOverview
Accurately tracking people in video enables applications in surveillance, traffic monitoring, and video conferencing. Recent multi-camera methods have helped to overcome some of the issues associated with object tracking, such as drift and occlusion, that arise in the single-camera case. However, the integration of multiple cameras introduces new challenges in terms of resource consumption (e.g., power, computing, and networking), and algorithm complexity. Moreover, while tracking accuracy tends to increase with the number of cameras, the potential also increases for poor measurements from individual cameras to negatively affect aggregate tracking estimates.
We introduce a method that balances reduced error with low algorithmic complexity by dynamically selecting the best subset of available cameras in which to perform active tracking. At each time frame, we evaluate the available cameras in terms of a ranking function (e.g. detection score) in order to choose the active cameras for the next frame.
Results show strong potential of our method to reduce tracking error with relatively few active cameras.
Video
Watch