Human Detection Based On Co-hog And Non-co-hog

In this paper, we mainly consider the problem of human detection in the static images. The goal of human detection is to determine the presence of humans and the positions of the humans. Human detection is crucial to image/video-based intelligent surveillance and thus is important for social security.Two key points of the human detection are feature extraction and the designing of the classifier. In this paper, we focus on the first one, i.e. feature extraction. One of the most successful features for human detection is Histograms of Oriented Gradients (HOG), which was proposed by Dalal et.al. in 2005. For improving the speed of the human detection and the efficiency of the classifier, Wang et.al. proposed to compute HOG features by orientation decomposition and convolution mask at International Conference on Computer Vision (ICCV) in 2009. For enhancing the expressiveness and discernment of HOG, Watanabe et. al. proposed Co-occurrence Histograms of Oriented Gradients (Co-HOG) in 2009.In this paper, based on the works above, we proposed three novel methods for improving the performance of human detection which are listed as follows:1) To improve the performance of the method proposed by Wang et. al., we proposed to employ mean filter for smoothing the gradients. Compared to Wang's filter which is linear to distance, the proposed method results in better detection performance.2) To further improve the performance of the method proposed by Wang et. al, proposed to enlarge the size of filter from 7×7 to 9×9.3) Existing Co-HOG uses merely the information of the direction of the gradient and thus the information of the gradient magnitude is neglected. To make full use of the information of the gradients, we propose to use not only the information of the direction but also the magnitude of the gradients.To sum up, we have implemented several classical human detection methods and proposed several novel methods to improve the performance of the human detection systems.