An end-to-end gait recognition system for covariate conditions using custom kernel CNN.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Gait recognition identifies individuals by their walking patterns, useful for non-intrusive surveillance.
• Covariate conditions like clothing significantly hinder accurate gait recognition.
• Current methods struggle to maintain performance under varying environmental and appearance factors.

Purpose of the Study:

• To propose a novel deep-learning framework to address covariate challenges in gait recognition.
• To improve the robustness and accuracy of gait identification systems in real-world scenarios.
• To develop an automated system that negates manual feature engineering for gait analysis.

Main Methods:

• A deep-learning framework was developed to identify and exclude regions affected by covariates during gait analysis.
• Customized kernels and feature extraction focused on dynamic, covariate-unaffected regions.
• A Convolutional Neural Network (CNN) was employed for feature learning and individual recognition from proposed regions.
• An end-to-end system was designed, integrating region proposal and feature extraction.

Main Results:

• The proposed method achieved 90% accuracy for gait recognition with subjects wearing bags.
• Accuracy reached 58% for subjects wearing coats, indicating robustness to clothing variations.
• High accuracy was observed for different walking speeds: 94% for fast and 96% for slow paces.
• Performance showed improvement over existing deep learning methods on CASIA datasets A and C.

Conclusions:

• The novel deep-learning framework effectively mitigates the impact of covariate conditions on gait recognition.
• The approach enhances the reliability of gait-based identification in practical surveillance applications.
• The automated system demonstrates superior performance compared to previous methods, particularly under challenging conditions.