HOG Features on the GPU

Histogram of Oriented Gradients (HOG) features are a trending topic in object detection literature. HOG features are a robust way of describing local object appearances and shapes by their distribution of intensity gradients or edge directions, and have been used successfully as a low level feature in a number of object recognition tasks. Human faces are generally considered interesting and important to detect in many applications such as surveillance, recognition systems, biomedical, and video. HOG descriptors have been shown to significantly outperform existing feature sets for human detection, but at the expense of greater computational overhead than other well known Haar-like rectangular features. We propose to accelerate the HOG feature descriptor computation by exploiting GPU parallelism so that existing frameworks for real-time detection can make use of a more robust feature.