Binary code-based Human Detection

HOG features are effective for object detection, but their focus on local regions makes them highdimensional features. To reduce the memory required for the HOG features, this paper proposes a new feature, R-HOG, which creates binary codes from the HOG features extracted from two local regions. This approach enables the created binary codes to reflect the relationships between local regions. Converting feature values to binary, however, results in the loss of much information included in the features. In response to this problem, we have been focusing on“ quantization residual”information that is lost at this time. In this study, we introduce a transition likelihood model into the classifier based on two ideas using quantization residuals to consider the possibility that a binary code observed from the image will make a transition to another binary codes. This enables classification that takes into account all binary codes including the originally desired binary codes even if an observed binary code differs from the truly desired binary codes due to some sort of effect from another binary codes. Experimental results show that a classifier equipped with transition prediction based on quantization residuals as proposed here achieves high-accuracy human detection compared to the same classifier without transition prediction.