Automatic understanding of video content is a problem which grows in importance every day. Video understanding algorithms require accuracy, robustness, speed, and scalability. Accuracy generates user confidence in usage. Robustness enables greater autonomy and reduced human intervention. Applications such as navigation and mapping demand real-time performance. Scalability is also important for maintaining high speed while expanding capacity to multiple users and sensors. In this thesis, I propose a "bag-of-phrases" model to improve the accuracy and robustness of the popular "bag-of-words" models. This model applies a "geometric grammar" to add structural constraints to the unordered "bag-of-words." I incorporate this model into an architecture which combines an object recognizer, a tracker, and a geolocation module. This architecture has the ability to use the complementarity of its components to compensate for its weaknesses. This allows for improvements in accuracy, robustness, and speed. Subsequently, I introduce VICTORIOUS, a fast implementation of the proposed architecture. Evaluation on computer-generated data as well as Caltech-101 indicate that this implementation is accurate, robust, and capable of performing in real time on current generation hardware. This implementation, together with the "bag-of-phrases" model and integrated architecture, forms a step towards meeting the requirements for an accurate, robust, real-time vision system.
展开▼
