Human attention and intent analysis using robust visual cues in a Bayesian framework

Human-computer interaction is beginning to permeate all aspects of our lives. As we develop more and more interactive computer systems, it is important to develop proper tools and frameworks for making this interaction more efficient. These types of interactive systems can also increase the safety of everyday tasks. We propose a human attention and intent analysis system based on a probabilistic framework using visual cues to help increase the efficiency and safety of everyday tasks. We will specifically look at the driving task. By combining cues about the vehicle interior and driver, the vehicle state, and the vehicle surround, we can make estimates of the driver's focus of attention and intent. Research contributions will be made in the areas of head pose and facial affect analysis, lane detection and tracking, fusing multiple cues to generate estimates of both attention and intent, and overall system integration. A Bayesian framework allows us to effectively combine multiple modalities of cues from visual information as well as other sensors to generate estimates which take into account the uncertainty of both the observations and the underlying process as well as prior knowledge about the parameters we are estimating. This can help us assess critical situations and feedback information faster and more efficiently than systems that do not take into account the driver's attention or intent. We will also show statistical results demonstrating the accuracy of such a system in real-world conditions using data collected from a 28 different drivers.