Automatic human gait imitation and recognition in 3D from monocular video with an uncalibrated camera (Q1954913)

Summary: A framework of imitating real human gait in 3D from monocular video of an uncalibrated camera directly and automatically is proposed. It firstly combines polygon-approximation with deformable template-matching, using knowledge of human anatomy to achieve the characteristics including static and dynamic parameters of real human gait. Then, these characteristics are processed in regularization and normalization. Finally, they are imposed on a 3D human motion model with prior constrains and universal gait knowledge to realize imitating human gait. In recognition based on this human gait imitation, firstly, the dimensionality of time-sequences corresponding to motion curves is reduced by NPE. Then, we use the essential features acquired from human gait imitation as input and integrate HCRF with SVM as a whole classifier, realizing identification recognition on human gait. In associated experiment, this imitation framework is robust for the object's clothes and backpacks to a certain extent. It does not need any manual assist and any camera model information. And it is fitting for straight indoors and the viewing angle for target is between \(60^\circ\) and \(120^\circ\). In recognition testing, this kind of integrated classifier HCRF/SVM has comparatively higher recognition rate than the sole HCRF, SVM and typical baseline method.