Nikos Papanikolopoulos, Professor, Computer Science and Engineering
Automatic extraction of events from video sequences has important applications for a variety of ITS problems, including scene monitoring, traffic data collection, and intersection monitoring. When deploying a system that recognizes events automatically from video sequences, two important things to consider are the real-time analysis of the video sequences and the short time required for learning the different classes of events in a scene. A related requirement that is often ignored is the limited reliance of the learning system on user-provided knowledge. In this project, researchers developed an innovative technique for detecting the different events in video sequences through a semi-supervised learning method. More concretely, the events are recognized from the tracked trajectories of the targets in the scene, which in turn are represented as a collection of actions or strings. By parsing these strings as reversible context-free grammars, the researchers detected and classified the different events. Learning consists of extracting the relevant grammar for each class of events from the data. To accomplish the learning goal, the system makes use of a small number of trajectories corresponding to each class as provided by a user to obtain a preliminary model of the grammar. Using this model, the system iteratively refines the grammar from new trajectory data obtained directly from the scene. Given that the system requires only a very small number of labeled trajectories and can iteratively learn from the observed data, the system is easily portable to new scenes with little system initialization from the user.