Facial detection and recognition are among the most heavily researched fields of computer vision and image processing. However, the computation necessary for most facial processing tasks has historically made it unfit for real-time applications. The constant pace of technological progress has made current computers powerful enough to perform near-real-time image processing and light enough to be carried as wearable computing systems. Facial detection within an augmented reality framework has myriad applications, including potential uses for law enforcement, medical personnel, and patients with post-traumatic or degenerative memory loss or visual impairments. Although the hardware is now available, few portable or wearable computing systems exist that can localize and identify individuals for real-time or near-real-time augmented reality.
The author presents a system design and implementation that performs robust facial detection and recognition robust to variations in lighting, pose, and scale. Scouter combines a commodity netbook computer, a high-resolution webcam, and display glasses into a light and powerful wearable computing system platform for real-time augmented reality and near-real-time facial processing. A convolutional neural network performs precise facial localization, a Haar cascade object detector is used for facial feature registration, and a Fisherface implementation recognizes size-normalized faces. A novel multiscale voting and overlap removal algorithm is presented to boost face localization accuracy; a failure-resilient normalization method is detailed that can perform rotation and scale normalization on faces with occluded or undetectable facial features. The development, implementation, and positive performance results of this system are discussed at length.