Histograms of Oriented Gradients

A useful question to ask of an image is whether it contains one or more instances of a certain object: a person, a face, a car, and so forth. Algorithms that answer this question are called object detectors. They often work by asking the same question in turn of every possible rectangle in the image that might possibly tightly bound one of the instances of interest. For instance, when detecting a pedestrian walking by a camera, one could start with a window that is 128 pixels tall and 64 pixels wide: The 2 to 1 aspect ratio of such a rectangle is a rough compromise between the aspect ratio of a person viewed from the front and one viewed from the side with legs fully extended during a step. The actual size of the window—using multiples of 64 is a mere convenience—reflects the assumption that if the image of the person is significantly smaller than that then resolution might be insufficient to detect it reliably, so it is not even worth trying. The person could of course occupy a bigger part of the image. Because of this, one analyzes not only the original image, but also those in a pyramid of images: A 128 by 64 rectangle at a coarser level of the pyramid corresponds to a larger rectangle in the original image (see Figure 1). The object detector then slides such a window over every image of the pyramid, perhaps in increments of a few pixels—the increment is called the stride of the detector—and computes a feature at every position, that is, a vector of numbers that describes the window’s contents. The feature is fed to the classifier, which tells whether the window contains a pedestrian. The answer is likely to be positive for a set of windows that overlap the person’s image, so the detector then chooses a single pixel to represent each connected set of positive classifier outputs. This note describes one way to construct a feature vector from a fixed-size window [1]. The feature is specifically tuned to pedestrians. This is because humans are important subjects in imagery, and is therefore natural that the detection of humans in still images and video has drawn much attention in the literature.