- The paper introduces Pose-Based Voting to robustly predict gender by conducting frame-by-frame analysis with Elliptic Fourier Descriptors and Row-Column Summation vectors.
- It employs Genetic Template Segmentation that uses genetic algorithms to identify optimal gait template regions, thereby mitigating challenges from covariate factors.
- The research advances authentication through Multiperson Signature Mapping and Bayesian Thresholding, significantly reducing error rates and enhancing system reliability.
Robust Analytics for Video-Based Gait Biometrics: An Overview
The doctoral thesis by Ebenezer R.H.P. Isaac presents advanced methodologies for video-based gait biometrics, addressing both soft and hard biometric challenges. Here's a formal summary of the research work and its contributions to the field of gait analysis:
The thesis provides four major contributions to gait biometrics. Initially, the development of a novel method for gait-based gender recognition known as Pose-Based Voting (PBV) is introduced. Unlike traditional methods that rely heavily on the complete gait cycle, PBV makes frame-by-frame predictions, aggregating these through a voting mechanism, which improves robustness against partial occlusion. The approach utilizes Elliptic Fourier Descriptors (EFD) and Row-Column Summation (RCS) vectors for feature extraction. The RCS method notably outperforms state-of-the-art techniques, achieving ideal gender classification accuracy within the CASIA-B dataset.
Furthermore, the research introduces Genetic Template Segmentation (GTS) for optimizing template-based gait recognition. This technique leverages a genetic algorithm to automatically determine the optimal regions of gait templates, thereby enhancing performance by concentrating on areas less affected by covariate factors such as clothing and carrying conditions. GTS is validated across multiple view angles, demonstrating superior results compared to conventional methods.
For gait authentication, the Multiperson Signature Mapping (MSM) paradigm is proposed to overcome limitations associated with traditional Euclidean distance-based thresholds. MSM utilizes a recognition framework where authentication is contingent on matching the predicted identity with the claimed identity, thereby effectively lowering the False Accept Rate (FAR) as a function of system population.
Addressing MSM's deficiencies in small system populations, Isaac advances Bayesian Thresholding (BT) as a superior methodology. This framework employs Bayesian posterior probabilities as authentication thresholds, exhibiting increased sensitivity and reliability. BT significantly reduces the Average Error Rate (AER) across all templates, offering enhanced performance relative to MSM.
In conclusion, Isaac's research substantially advances the understanding and application of gait biometrics, delivering sophisticated techniques for gender recognition and identification, and contributing innovative authentication paradigms that promise scalability and resilience against spoofing. Future work could extend these frameworks to incorporate diverse feature sets and apply them across alternative modalities, presenting an opportunity to enrich verification systems across broader applications.