Statistical gabor-based gait recognition using region-level analysis

Gait recognition has become a popular research problem gaining importance for human identification based on walking style. It has emerged as an attractive research problem due to possessing several desirable merits unlike other biometrics. However, most of the existing gait recognition methods that involve Gabor-based filters suffer from the curse of dimensionality, even with the use of a dimensionality reduction technique. This adds more computational and storage burdens and may cause difficulties to identify subjects with a high degree of confidence. In this paper a statistical gait recognition approach is proposed based on the analysis of overlapping Gabor-based regions. The Gait Energy Image (GEI) is first constructed from the gait sequence as a spatio-temporal summary. Then, the GEI image is convolved with a Gabor filter bank of 8 different orientations and 5 different scales. A statistical analysis is then applied to extract discriminative gait features from multi-overlapped Gabor-based regions. Consecutively, an SVM classifier is applied to measure the gait similarity and identify the subject. Comprehensive experiments are carried out to evaluate the proposed approach and compare it to existing approaches. The results have shown that promising performance can be achieved with the proposed approach under a variety of scenarios.