A Real-Time People-Counting System Based on Improved HOG

This paper designs a real-time people-counting system based on improved histogram of oriented gradients (HOG). This paper has taken three measures to improve the computation speed of the detection algorithm: detecting human bodies with extraction of moving regions; determining the region of interest (ROI) and then the dimension of feature vector is reduced from original 3780 to 1764; using the graphic processing unit (GPU) programming method. This paper collects human samples from the actual scene and then puts it together with the INRIAPerson library as samples for training classifiers, which improves the accuracy of detection. Meanwhile, in order to get better continuity of human detection, three classifiers are designed: the whole body classifier (W-Classifier), the head classifier (H-Classifier), and the leg classifier (L-Classifier). The average detection time of the proposed system is 52ms per frame with the resolution of 320 x 240. Therefore, it meets the real-time requirements, and the correct detection rate reaches nearly 90%.