Localized Hough Transform for Human Detection

This paper presents a method that uses Hough transform in the localized area for robust visual object recognition. We focus on the human detection problem at different resolutions. The proposed method takes into account orientation information along with spatial information, which is ignored by some dense image descriptors such as the histogram of oriented gradients (HOG). With our method, we capture the linear structure in the image, which renders more discriminative features for object detection than the existing methods. We also discuss in detail as to how each parameter of the localized Hough transform (LHT) affects the robustness of object detection, with experiments done on human detection. Implications of choosing image descriptors versus learning based methods for human detection at various scales are also discussed in this paper, comparing the current state-of-the-art methods for human detection including the convolutional neural networks, the histogram of oriented gradients method and our method. Our experimental results show that the proposed method has the best performance.