Human head detection using Histograms of Oriented optical flow in low quality videos with occlusion

Video detection and tracking of humans in large events is a non-trivial task. The tracked information is useful in calculating crowd behavior parameters such as density, speeds and paths that are critical in making automated decision about crowd behavior. During occluded scenarios, it is very common to have only head and shoulder visible in the videos and the rest of the body is blocked by other people or objects. Although there is wealth of information in literature, tracking in occluded high density environments is not well addressed. In this paper, we aim to detect head and shoulder of people by hybrid motion and visual features in a low quality video with occlusion and cluttered environment. We propose an improved version of histogram of oriented optical flow (HOOF) called integral HOOF. The HOOF is not only shown to be a discriminative feature in low quality video, but also an alternative way to segment moving objects in a video. HOOF and Histogram of Oriented Gradient (HOOG) features are extracted from a real world surveillance video, and Support Vector Machine classifier is trained to detect heads and shoulders. Experimental result shows the HOOF has high precision, recall and accuracy on detecting heads and shoulders.