Due to the complexity of human motion tracking in video, especially in complex background, there exist some problems in the region of human lower extremity recognition and tracking algorithm. Based on traditional Kalman filtering, a discrete Kalman filter body recognition algorithm is proposed based on variable measurement covariance. By measuring the covariance initialization, new covariance estimators are calculated from the observation data with the recursive method, and tracked by the discrete Kalman filter. The method show good tracking effect and performance when applying to the video image in real, and to distinguishing and tracking parts of upper limb, lower limb of the human body. Compared with traditional Kalman filter algorithm, the proposed method does not lose the tracking targets, the tracking speed is moderate maintaining consistent to human walking speed, which basically is 1.5 m/s. This algorithm is especially suitable for the video tracking and analysis for human behavior.