Automatic age and gaze estimation under uncontrolled conditions

Can the computer learn a person’s age just from analyzing facial images obtained by a simple webcam? If so, you can automatically determine if a person has the right age to buy beer (shops), to use a credit card (retail), and can enter bars (entertainment). Can the computer automatically recognize where people are looking at? This is very useful if you want to automatically determine the viewing of a person when looking at websites, products in shopping malls, and the road when driving a car. This thesis addresses these two questions by providing algorithms for age prediction and eye gaze estimation. Different from previous approaches, the focus is to provide (practical) solutions for real-world applications. Automatic age estimation may be hindered by many different factors. One of the main problems is facial expressions as they create expression-dependent face wrinkles confusing the real age wrinkles. Moreover, poor imaging conditions may introduce noise that may affect the accuracy of the estimated age. We go beyond the standard scenarios to address these cases. Typical gaze estimation systems require calibration which involves following explicit instructions by the user such as tracing a moving dot on the screen. Such prerequisite is not feasible for some practical applications such as shopping malls to determine where the people are looking at. In this thesis, algorithms are developed to estimate the gaze points without the need for a calibration by the user.