Covariate factor mitigation techniques for robust human gait recognition
Imagining trying to recognise a person from CCTV images. Traditional biometrics
such as facial recognition is rendered ineffective should a suspect's face be
concealed. Gait can identify a person based on their unique walking manner and
posture. As a biometric, gait is favourable as the capture requires no cooperation,
consent or contact from the suspect.
Before gait recognition can be deployed during real world surveillance, robustness
to common real world covariate factors must be established during controlled validation
datasets. Our research focuses on boosting the robustness to covariate factors which
affect the natural gait appearance and motion; examples include clothing, shoes,
carrying a bag, elapsed time between capture (months) and complex couples thereof.
Our most recent work exploits the novel exploitation of popular single compact 2D gait representations and model-free skeleton-based representations to yield the Skeleton Variance Image (SVIM). Through space- and time-normalisation, a gait sequence is condensed into a single compact 2D gait representations which is highly beneficial for memory and computational processing costs. Skeleton-based representations are sparsely employed in gait recognition due to the inherent boundary perturbation sensitivity occurring from the imperfect extraction of silhouettes. By employing a smooth distance function generated by the Poisson equation, we generate robust skeletons capable of differentiating between the natural gait and covariate factors - overcoming a significant limitation of gait recognition. To further combat the effects of covariate factors, motion features are extracted for their saliency and consistency over time compared to their appearance feature counterparts. Our robust gait recognition approach achieves state of the art results when validated on the standardised publicly available TUM GAID dataset, shown below:
Tenika Whytock Dr. Neil M Robertson