Real-time face detection during the night

Abstract

To address the challenges in unconstrained real scenes face detection during the night, we propose a real-time night face detector. First, we use histogram equalization methods to process the input image of night to increase its contrast. Second, the image features called Promotion Normalized Pixel Difference [3] (PRO-NPD) is proposed. PRO-NPD features are computed as the ratio of difference to sum between two pixel values. Third, we use a deep quadratic tree to learn the optimal subset of PRO-NPD features and their combinations. We construct a Face Detection at Night in Surveillance (FDNS) database with 741 images in actual surveillance environment with various lighting. Experiment performed well in both accuracy and speed. We achieved 30 fps high-speed face detection in a robot with embedded ARM Cortex A17 quad-core 1.4GHz CPU.