Behavior histograms for action recognition and human detection

Abstract. This paper presents an approach for human detection and simultaneous behavior recognition from images and image sequences. An action representation is derived by applying a clustering algorithm to se-quences of Histogram of Oriented Gradient (HOG) descriptors of human motion images. For novel image sequences, we first detect the human by matching extracted descriptors with the prototypical action primitives. Given a sequence of assigned action primitives, we can build a histogram from observed motion. Thus, behavior can be classified by means of his-togram comparison, interpreting behavior recognition as a problem of statistical sequence analysis. Results on publicly available benchmark-data show a high accuracy for action recognition.