Video Monitoring System: Counting People by Tracking

We present a video monitoring system to count the number of people in an open area such as an airport, a bus station, or a shopping mall using a single camera. Our system automatically infers the number of people based on a novel multi-people tracker. The tracking framework is formulated as a data association problem, in which the people are detected in every frame and then all of the detection responses are associated to form the right track for each person. The detection step is carried out by an efficient state-of-the-art method while the data association is done by using a hierarchical framework, in which the detection responses form short tracklets, the short tracklets form longer ones, and so on. After that the number of people is counted based on the number of tracklets, from which the number of entered and exited people is also addressed. Beyond the scope of this system, we also aim to assign the right person into the right track instead of counting them only by proposing a novel appearance model to help re-identify a person, especially after heavy occlusion. This proposed solution is equivalent to addressing the problem of ID switching in multi-target tracking. The system is tested in both indoor and outdoor environments and in public datasets together with our new dataset. We also compare different components in our framework and other state- of-the-art trackers.