Non-background HOG for pedestrian video detection

Histogram of Oriented Gradient (HOG) features are proved to be very effective for pedestrian detection in static image. However, most of the background information is wasted when the features are used to detect human in video. Especially in complex environment, the non-eliminated background gradient will affect the detection results. To improve the overall detection performance, a new feature named Non-background HOG is proposed which created a cell map using GMM for the procedure of image gradient calculation in HOG algorithm. This new algorithm not only is capable of reducing the influence of background gradient, but also speeds up the extraction running time. Evaluation experiment demonstrated that the non-background HOG algorithm gives a better performance than classic HOG in pedestrian video detection.