Human Identification Based on Gait
Human Identification Based on Gait is structured to meet the needs of professionals in industry, as well as advanced-level students in computer science.
Human Identification Based on Gait
" 7 Further Gait Developments (p. 135-136)
7.1 View Invariant Gait Recognition
Human gait characteristics are best analyzed when presented in its canonical view (side view). The dynamics behind an individuals gait, the quasi-periodicity of an individuals gait etc. which contribute significantly towards establishing his/her identity are more evident when the canonical view of human gait is presented . Thus, the availability of such a canonical view is crucial to the success of many gait recognition systems.
We observe that in surveillance applications a non-canonical view of an individuals gait is captured more often and this necessitates developing gait recognition algorithms that are view invariant. When individuals walk at an angle oblique to the camera, the performance of gait recognition systems suffers due to the gradual change in the individuals height and stride length as captured by the camera. One approach could be to build a 3D model of an individual and extracting viewinvariant features that best characterize the individuals gait. But such an approach typically based on Structure from Motion (SfM) and stereo reconstruction techniques suffer from many shortcomings.
Shakhnarovich et.al.  compute a visual hull from a set of monocular views and render virtual canonical views for tracking and recognition. They extract moments based image features from the silhouettes of the synthesized video to perform gait recognition. Naturally, a model of walking is implicitly invariant to viewpoint for a small change in viewing angle. On the structural side, Bobick et al.  developed a gait recognition technique that recovers static body and stride parameters of subjects from their walking sequences. The parameters are evaluated by means of a confusion metric which predicts the effectiveness of the feature vector in identifying an individual. They define a mapping function across different views and use the same to perform gait recognition.
The stride parameters that were extracted from each subject were shown to be more resilient to different viewing directions. Naturally, BenAbdelkaders stride and cadence structural approach equally has viewpoint invariant properties . On the pure modeling side, Carter et al. showed  that variations in gait data could be corrected after the data had been gathered, given that the trajectory is known. This was achieved by considerations of geometry and use of a simple pendulum modeling the thigh, with correction by rotation of the thigh swing axis. A Fourier analysis showed that viewpoint correction had been achieved.
The approach was then reformulated to provide a pose invariant biometric signature which did not require knowledge of the subjects trajectory. Spencer et al. was to extend these notions  and developed a geometric correction to the measurement of the hip rotation angle, based on known orientation to the camera. Results on synthesized data showed that simple pose correction for geometric targets generalizes well for objects on the optical axis. Later, these techniques were refined for analysis on walking subjects, and showed that the approach can work well, given that target features can be extracted and tracked with success ."