Extracting movement, posture, and temporal style features from human motion

Small variations in biological motion responsible for perception of characteristics, styles, or affects of the person performing the actions, are referred to as secondary features. This paper presents a novel method for separating and extracting spatiotemporal sets of secondary features from human motion data. The technique employs a dataset of sequences and identifies a corresponding neutral sequence through maximizing a similarity index based on correlation. Specific control points or temporal cues are then distributed through the input sequence. Distribution is carried out with the goal of maximizing an objective function successive to time warping. The optimized set of cues are used to reconstruct the neutral component of the signal using cubic splines. Accordingly, both spatial (movement and posture) and temporal secondary features are extracted from the stylistic input sequence. To illustrate one of the possible applications of the proposed technique, style translation is carried out. We illustrate that our proposed system can be used to extract various classes of secondary features from different actions such as walking, jumping, and running.