Effective cue integration for fast and robust face detection in videos

We propose a novel framework using multimodal cues effectively and efficiently for a fast and robust detection of human faces in video images. Proper integration of cues may allow for both acceleration of the process and reducing the false positive errors while retaining a high rate of detection. In our algorithm, we integrated three salient cues based on motion, color, and appearance which are complementary features characterizing the face. For efficient integration of the cues, the sequence of images is processed in two stages: pre-attentive stage and post-attentive stage. In particular, a decision maker in pre-attentive stage is proposed to minimize the processing time, which switches the route of the information flow of the process according to the current scene properties. With proper sequential usage of the three multimodal cues depending on scene properties, the proposed method outperforms conventional approaches in terms of detection rate, false positives, and computational cost by reducing markedly the search space and search scale of the appearance-based face detector with low computational cost.