The paper presents an automatic initialization procedure for visual tracking of human motion. Instead of relying merely on low-level image features to give a single estimate of the initial human posture, the system seeks to find a set of sam-ples that carries multiple hypotheses of the pose. By accu-mulating different image cues in the first 3-15 consecutive frames and combining dynamic information regarding hu-man motion, the system builds a human body model for the person to be tracked from a video sequence and produces a sample set as an estimate of the posterior distribution of the initial posture. The sample set provides a good starting point for tracking with sequential Monte Carlo methods. 1.