UC Riverside

The traditional approaches for pedestrian tracking are only focused on pure frame-based vision features. But the temporal and spatial relationship between pedestrians in real-world coordinates is somewhat ignored. My research is focused on how to apply spatial and temporal constraints in pedestrian tracking, and further improve the tracking performance.

In order to achieve this goal, one way is to take the advantages of crowd simulation approaches from computer graphics. Before that, the first step is to train the parameters of different crowd simulators based on pedestrian data from real videos, so that the simulated trajectories can be as realistic as possible. Then we built up pedestrian tracking systems with crowd simulators integrated. The output of crowd simulation is treated as extra prediction when traditional vision-based tracking approaches are used. In addition, to further investigate the influence of crowd simulation, two simulators with different strategies are integrated and the outputs from the two tracking systems are compared.

However, pedestrians under different situations may need different combination of crowd simulation model and its parameter settings for the best simulation. Therefore, it would be important that the crowd simulator to be more flexible for different pedestrians. Accordingly, a new context-aware crowd simulator is proposed and integrated to our tracking system as well. This new crowd simulator has the ability to switch between different simulation models as well as their different parameter settings, based on the context feature extracted for different pedestrians.

Another type of spatial and temporal information that we have explored is the group structure. Since groups can be frequently observed in crowded scenes, and the pedestrians within a same group usually have adjacent locations and similar velocities, it is reasonable to track pedestrians in a group together. A structure preserving object tracking approach that is able to use group information is extended for pedestrian tracking in a multi-camera video network.

For all tracking approaches, the performance is evaluated by multiple object tracking precision and accuracy metrics. The experimental results demonstrated that these types of spatial and temporal constraints among pedestrians boost the tracking performance significantly, especially for a crowded scene.