A framework for face detection on Central Catadioptric Systems

This paper presents a method of adaptation of the well-known Viola and Jones' face detector to Central Catadioptric Systems. The performance of this detector is well-known and has been well studied. Our purpose is not to evaluate the intrinsic limits of the Viola and Jones' algorithm but to evaluate the impact of central projection systems on the algorithm's behavior. 360-degree field of view sensors offer interesting capabilities for Human-Robot Interface. They do not require users to be in a specific part of the robot's surrounding environment. Most of the time, such interfaces imply face detection. Unified models can be used to calibrate central projection systems resulting in a single set of parameters. In order to apply perspective-like detection techniques, we apply rectification techniques to transform wrapped images into perspective panoramic images. We demonstrate that spherical projection gives better results than cylindrical projection. We introduce a model of a person's location with respect to the sensor. Spatial limitations and the impact of the interpolation method during the unwrapping procedure are also discussed.