Multi-View Gait Image Generation for Cross-View Gait Recognition.

Area of Science:

• Biometrics
• Computer Vision
• Artificial Intelligence

Background:

• Gait recognition offers non-contact, long-distance identity verification, surpassing traditional methods like face or fingerprint recognition.
• Cross-view gait recognition is challenging due to significant variations in gait appearance caused by different viewing angles.
• Current deep learning methods for cross-view gait recognition are hindered by insufficient gait samples across multiple views.

Purpose of the Study:

• To address the data scarcity issue in cross-view gait recognition by generating synthetic gait samples.
• To improve the performance of deep learning models for cross-view gait recognition through data augmentation.
• To introduce a novel generative adversarial network for multi-view gait data synthesis.

Main Methods:

• Development of a Multi-view Gait Generative Adversarial Network (MvGGAN) capable of generating diverse gait samples.
• Training a single generator to handle all view pairs within single or multiple datasets.
• Implementation of domain alignment using projected maximum mean discrepancy to mitigate distribution divergence.

Main Results:

• The MvGGAN successfully generated realistic fake gait samples that augmented existing datasets.
• Experimental results on CASIA-B and OUMVLP datasets demonstrated significant performance improvements in cross-view gait recognition.
• The proposed method enhanced state-of-the-art cross-view gait recognition across both single-dataset and cross-dataset evaluations.

Conclusions:

• The MvGGAN effectively generates gait samples, overcoming the limitations of data scarcity in cross-view recognition.
• The synthetic data generated by MvGGAN significantly boosts the performance of deep learning-based cross-view gait recognition systems.
• This approach offers a viable solution for improving the robustness and accuracy of gait recognition systems in real-world applications.