Histogram of Oriented Gradients Algorithm

The histogram of oriented gradients, known as HOG, is a feature-description algorithm that summarizes the shape and appearance of an object in an image by counting how often gradients point in each direction within small regions of the image. It divides the image into a grid of small cells, computes the direction and strength of the local intensity gradient at each pixel, and builds a histogram of gradient directions for each cell; the histograms from a block of neighboring cells are then concatenated and normalized against local contrast, which makes the final descriptor considerably more robust to changes in lighting than the raw gradients alone would be. Navneet Dalal and Bill Triggs presented the technique in 2005, developing it specifically to detect pedestrians in photographs, and it went on to become a standard feature descriptor for object detection tasks generally, in the years before deep learning based detectors became dominant.