Nyan Bo Bo, Matthew N. Dailey, and Bunyarit Uyyanonvara

pp. 157 - 165

Abstract

Robust real-time hand detection and tracking in video sequences would enable many applications in areas as diverse as human-computer interaction, robotics, security and surveillance, and sign language-based systems. In this paper, we introduce a new approach for detecting human hands that works on single, cluttered, low-resolution images. Our prototype system, which is primarily intended for security applications in which the images are noisy and low-resolution, is able to detect hands as small as 24×24 pixels in cluttered scenes. The system uses grayscale appearance information to classify image sub-windows as either
containing or not containing a human hand very rapidly at the cost of a high false positive rate. To improve on the false positive rate of the main classifier without affecting its detection rate, we introduce a post-processor system that utilizes the geometric properties of skin color blobs. When we test our detector on a test image set containing 106 hands, 92 of those hands are detected (86.8% detection rate), with an average false positive rate of 1.19 false positive detections per image. The rapid detection speed, the high detection rate of 86.8%, and the low false positive rate together ensure that our system is useable as the main detector in a diverse variety of applications requiring robust hand detection and tracking in low-resolution, cluttered scenes.