Fast human detection based on enhanced variable size HOG features

In this paper, we proposed an enhanced variable size HOG feature based on the boosting framework. The proposed feature utilizes the information which is ignored in quantization gradient orientation that only using one orientation to encode each pixel. Furthermore, we utilized a fixed Gaussian template to convolve with the integral orientation histograms in order to interpolate the weight of each pixel from its surroundings. Either of the two steps have an important effect on the discriminative ability of HOG feature which leads to increase the detection rate. Soft cascade framework is utilized to train our final human detector. The experiment result based on INRIA database shows that our proposed feature improves the detection rate about 5% at the false positive per window rate of 10^-4 compared to the original feature.