GaitMPL: Gait Recognition With Memory-Augmented Progressive Learning.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Gait recognition identifies individuals by their walking patterns, but faces challenges from appearance variations (intra-class diversity) and similar gaits (inter-class similarity).
• Existing methods struggle with these 'hard samples,' limiting recognition accuracy, especially at distances and under varying conditions.

Purpose of the Study:

• To address the hard sample issue in gait recognition.
• To propose a novel Memory-augmented Progressive Learning network (GaitMPL) for enhanced pedestrian identification.

Main Methods:

• Developed GaitMPL, incorporating a Dynamic Reweighting Progressive Learning module (DRPL) for easy-to-hard sample learning.
• Integrated a Global Structure-Aligned Memory bank (GSAM) to model feature distributions for each individual ID.

Main Results:

• GaitMPL achieved state-of-the-art performance on the CASIA-B dataset, reaching 88.0% accuracy under the 'Clothing' condition and 93.3% on average.
• Demonstrated significant improvements over existing methods, outperforming them by at least 3.8% and 1.4% respectively on challenging and average conditions.

Conclusions:

• GaitMPL effectively solves the hard sample problem in gait recognition.
• The proposed network shows superior performance and robustness in pedestrian identification tasks.