EE Seminar: Gait Based Person Identification Using Motion Interchange Patterns
~~Speaker: Gil Freidlin,
M.Sc. student under the supervision of Prof. Shai Avidan
Wednesday, August 5, 2015 at 15:00
Room 011, Kitot Bldg., Faculty of Engineering
Gait Based Person Identification Using Motion Interchange Patterns
Abstract
Human gait is a valuable biometric characteristic which, among other biometric measures, has attracted much attention in recent years. Vision-based gait analysis for identification offers several advantages over other biometrics, as gait can be recognized from a distance, does not require cooperation or even awareness of the subject, and can work well even on low resolution videos as recorded by standard surveillance cameras.
Over the years many attempts have been made to design effective representations of human motion in video content. Commonly, most representation methods applied to gait recognition were based on silhouettes extraction, making their performance sensitive to variations in silhouette’s visual appearance caused by different walking conditions and to the silhouette’s quality, and less suitable for use in unconstrained environments. In the related task of action recognition, video representation aims to distinguish among human actions. Interestingly, some motion-based representations developed for action recognition and recently applied for gait recognition have achieved promising results.
Encouraged by this line of work, we propose a new approach for gait representation by adopting the Motion Interchange Patterns (MIP) framework, that was successfully employed in action recognition. This effective framework encodes motion patterns by capturing local changes in motion directions, to extract local motion-based features directly from the video. A bag-of-words scheme is used to construct the gait representation of a recorded walking person. Our framework does not rely on silhouettes commonly used in gait recognition, and benefits from the capability of MIP encoding to model real world videos. We also propose two adaptations of MIP that are used to enrich the motion-based gait representation.
We empirically demonstrate the effectiveness of this modeling of human gait on several commonly used gait recognition datasets. The proposed approach is shown to perform well for various walking conditions, especially showing robustness in the challenging case of a difference in carrying conditions (walking while carrying a bag), wherein it consistently outperformed other reported methods.
