Learning Models of Activities

The goal of this research is to automatically learn semantic scene models from long term observations of object activities in the scene. The scene structures of interest include paths of vehicles and pedestrians, entry and exit regions, parking, drop-off, and pick-up places, etc. They are all related to different types of activities. The learnt semantic scene models provides both geometric and statistical description on when, where, and what types of activities occur. They can be applied to abnormal activity detection, activity description, object classification, and tracking, etc. as prior knowledge on the scene.