An adaptable architecture for human-robot visual interaction

Face recognition has received increasing attention during the past decade as one of the most promising applications of image analysis and processing. One emerging application field is Human-Machine Interaction involving robotic vision. For many applications in this field (including face identification and expression recognition) the precision of facial feature detection and the computational burden are both critical issues. This paper presents a completely tunable hybrid method for accurate face localization based on a quick-and-dirty preliminary detection followed by a 2D tracking. Our technique guarantees complete control over the performance/result quality ratio and can be successfully applied to intelligent robotic vision. We use our approach to design a Robotic Vision Architecture capable of selecting from a set of strategies to obtain the best results.