This paper presents a new approach based on temporal minimization for separation and extraction of high/low-energy variants embedded in human motion. A data set of over 6500 frames is used for training the proposed algorithm. Spatiotemporal cubic splines are employed for approximating the trajectories associated with walking sequences. The optimal numbers of control points required for synthesizing the neutral movements are calculated. We illustrate that by minimizing an error value with respect to the training data set and reconstructing the trajectories, the low and high-energy variants can be separated from the main gait and hence extracted.