Robust face detection at video frame rate based on edge orientation features

In this paper, we present a novel approach to real-time face detection in arbitrary grey images and video streams using a multi-stage detection pipeline. It is based on edge orientation matching (EOM) and subsequent candidate verification using a SNoW (Sparse Network of Winnows) classifier in the final stage. We have developed a simple and efficient EOM method for object modeling and matching using edge orientation information only. To make the system more robust, we verify face candidates obtained by the EOM with a view-based SNoW classifier which operates directly on the grey values. For further speedup, the pipelined matcher is combined with a coarse-to-fine grid scan of the image. On average, it thus requires less than 10% of all possible image locations to be processed by the detection pipeline. With this method, it takes 50 ms on an Athlon 1000-MHz PC to analyze an image of spatial size 384/spl times/288 pixels. Experimental results on a large database of 18,704 grey still images with 19,030 detectable frontal faces, including the M2VTS (Multi-Modal Verification for Teleservices and Security applications) frontal face database, the CMU (Carnegie Mellon University) test set of H.A. Rowley and S. Baluja (1996) and the newly released BioID (Biometric Identification) test set, show that the presented approach outperforms methods that use neural networks, Bayesian methods and support vector machines (SVMs) by far in terms of computational efficiency while showing comparable detection results. The overall recognition performance is above 96%.