Solutions for Model-Based Analysis of Human Gait

The analysis-by-synthesis concept is applied in markerless human gait analysis. Human locomotion is approximated by means of adaptive tracking with a 3D model that moves in exactly the same manner as the subject in front of the cameras. This paper focuses on two particular problems: (1) the inverse mapping of pixels from the synthetic image back to the surface of the 3D model, and (2) the acquisition of initial values for automatic initialization of the 3D model for subsequent reliable tracking. Some interesting initialization constraints arise when the analysis-by-synthesis concept is applied in medical human gait analysis. The moving subject is segmented with an improved dual difference technique, which uses the gradient norms of real camera images. The most important assumption is that human gait is almost completely periodic. This allows a much more robust approach whereby the keyframe animation technique serves to synthesize artificial motion patterns using approximately correct joint angles.