Silhouette-based gait analysis is a technique for identifying whether two pedestrian videos depict the same
person. However, when the pedestrian is captured close to the camera, projective distortion causes their shape to change
nonlinearly. This shape change can make it difficult to correctly determine if it is the same person, even if the camera viewing
angle with respect to the person is only slightly different. In this study, we proposed a gait analysis method in which the
shooting angle, which is a three-dimensional (3D) camera parameter, of each pedestrian video was first calibrated and then the
silhouette video for learning was computed (perspective projection or simulation) from a four-dimensional (3D + time) gait
database so that the view angle of the silhouette corresponds to that of each video. We examined the person identification rate
of the proposed gait analysis and the relationship between the degree of the projective distortion and the degree of
improvement in the individual identification accuracy using the proposed method. Consequently, the proposed method was
found to be effective when the projective distortion was strong, i.e., when the pedestrian is close to the camera.
View full abstract