A model for transition-based visuospatial pattern recognition

Abstract

In my research, I designed and implemented a system for learning and recognizing visual actions based on state transitions. I recorded three training videos of each of 16 actions (approach, bounce, carry, catch, collide, drop, fly over, follow, give, hit, jump, pick, push, put, take, throw), each lasting 10 seconds and 300 frames. After using a prototype system developed by Dr. Satyajit Rao for focus and actor recognition, actions are represented as qualitative state transitions, tied together to form tens of thousands of patterns, which are then available as action classifiers. The resulting system was able to build simple, intuitive classifiers that fit the training data perfectly.