Left-Luggage Detection using Bayesian Inference

This paper presents a system that incorporates low-level object tracking and high-level event inference to solve the video event recognition problem. First, the tracking module combines the results of block tracking and human tracking to provide the trajectory as well as the basic type (human or non-human) of each detected object. The trajectories are then mapped to 3D world coordinates, given the camera model. Useful features such as speed, direction and distance between objects are computed and used as evidence. Events are represented as hypotheses and recognized in a Bayesian inference framework. The proposed system has been successfully applied to many event recognition tasks in the real world environment. In particular, we show results of detecting the left-luggage event on the PETS 2006 dataset.