Three-dimensional Human Motion Tracking

Human motion analysis and tracking from video has been an important research topic in the fields of image processing and computer vision during the past few years. It plays an important role in smart surveillance system, virtual reality, model-based coding, performing gesture and event understanding, man-machine interface, etc. Generally the procedure of human motion tracking and analysis from a sequence of images involves three main stages: (1) human body segmentation in a complex scene; (2) human motion tracking and body structure re-construction; (3) motion analysis and action recognition. As the base of the human action recognition and understanding, human motion tracking and body structure reconstruction is the key of the procedure. However, most previous efforts are concentrated on monocular, two-dimensional (2D) tracking, which makes the acquisition of the three-dimensional (3D) coordinates of human's body parts very difficult and impossible to reconstruct the 3D human posture. In this paper, a key frame based method is put forward to track the motion of human upper limbs in walking and at last the sequence of 3D human skeleton is recovered.Two image sequences are captured by two cameras in the parallel axis stereo imaging system and key frames are extracted based on the changing trend of gravity center during human walking.In the key frames, the features of body parts (for example, the edge information of head-shoulder region and skin color of human hand) are used to locate the 2D position of human head, shoulders and hands. Then their 3D coordinates are computed under the parallel axis stereo according to the disparity between the left and right images. The other joints in upper limbs are found by the skeleton structure and multiple constraints in human motion.Between the key frames, suppose human walking activity as the superposition of an ideal basic motion with, for example, constant velocity and white noise and an unproved Kalman predictor is adopted to estimate the 3D position of upper limbs' joints.The last part of the procedure is 3D recovery. Since the 3D coordinates of upper limbs' joints both in and between the key frames have been acquired, the data are used to reconstruct the posture of upper limbs.In conclusion, the key frame based method proposed in this paper avoids the incremental error of motion prediction in a too long image sequence and gives precise results in tracking which can meet the demand of precision in many applications. This method is effective and novel to some extent. At last, the further suggestions for the improvement are discussed.