Key points are not available for this paper at this time.
Understanding human behaviors and generating human-like motions are key technologies for human-robot interaction, motion synthesis in computer animation, sport training, rehabilitation. The observation of human actions is fundamental in understanding and generating the motions. The technology of human motion capture has been developed, and marker-based motion capture systems are used in various research fields. However, the marker-based motion capture systems have several drawbacks of expensiveness, obtrusiveness, and operational complexity. Markerless motion capturing has potential to overcome these drawbacks. Recently a large amount of 3 dimensional human whole body motion data has been accumulated , and research on structuring the motion knowledge from these motion data has been made. The structuring of a large amount of motion data is expected to yield new approaches to recovering 3 dimensional motion from monocular images by reusing the accumulated motion data. This paper describes the motion database which memorizes 3 dimensional motion data, the sequences of 2 dimensional image features, and the relations between these data through stochastic models. The 3 dimensional motion corresponding to the input sequence of images can be retrieved efficiently by using the stochastic models.
Takano et al. (Sun,) studied this question.