Fast human detection using mi-sVM and a cascade of HOG-LBP features

This paper presents a human detection approach which can process images rapidly and detect the objects accurately. The features used in our system are the cascade of the HOG (Histograms of Oriented Gradients) and LBP (Local Binary Pattern). In order to achieve high recall at each stage of the cascade, we modify the mi-SVM (Support Vector Machine for multiple instance learning) to train the HOG and LBP features respectively. In this way, we implement a novel cascade-ofrejectors method to detect the human fast, while maintaining a similar accuracy reported in previous methods. Experimental results show our method can process frames at 5 to 10 frames per second, depending on the scanning density in the image.