The invention discloses an automatic robust three-dimensional face detection method. According to the method, three-dimensional face discrete point cloud having plenty of redundancy information and noise under different attitudes and expressions are taken as input, and pre-processing, and three stages including nose tip point positioning and face detection are mainly included; for the pre-processing stage, smooth processing on an inputted point cloud model is firstly carried out, secondly, face pre-segmentation for the three-dimensional model is carried out; for the nose tip point positioning stage, firstly, initial face rotation transformation is carried out, and a discrete contour point set of an original face is acquired, secondly, curve fitting and gradient analysis are carried out, and a most-prominent point is extracted from a nose contour as a nose tip point; for the face detection stage, the nose tip point is a ball center, an empirical value is taken as a radius to segment to obtain a face region, and attitude correction is further carried out. The method is advantaged in that the method is applicable to a large-scale three-dimensional face database, good robustness for the attitudes and the expressions is realized, a processing speed is fast, and accuracy is high.