Several works that use colour as the main cue for object detection often assume static colour models. However unless using a colour constancy approach, or working in well controlled environments, different lighting conditions can lead to high colour variability that may make the detector fail. This variation in lighting conditions is even more critical in robotic applications working in uncontrolled environments, that are supposed to move along different areas and rooms with different illuminations. In this paper, a technique for face detection using non-parametric colour model is presented, as well as a method for adapting the colour model to better fit the illumination conditions, and the specific colour of the face of the person being tracked.