Adaptive, Automated Real-time Event/Target Detection using Supervised Learning

Automatic event and activity recognition from stream video and video archive can greatly enhance the effectiveness and efficiency of motion imagery analyst by filtering out the most relevant and timely content from a huge amount of video data through query video clips. Building on our DARPA VIRAT program to detect various events, such as vehicle starting, stopping, turning, U-turning, person running and so on, we propose to improve novel event detection by leveraging existing event modules to build the new model through semi-supervised learning and label expansion. The proposed approach will allow the user to annotate just a few training samples, and it will automatically label additional data based on other, existing event models, thus improving the new model over one built just on the user provided samples. We also propose to explore a novel multi-way interaction descriptor based on disjoint information that will be able to describe multi-entity events, and will improve the performance of single entity events by jointly using multiple features.