Learning Efficient Spatial-Temporal Gait Features with Deep Learning for Human Identification.

Area of Science:

• Biometrics
• Artificial Intelligence
• Computational Biology

Background:

• Human gait is a unique biometric for identification, offering remote and robust features.
• Existing methods struggle with gait's intra-class variations and inter-class similarities.
• Advancements in AI and bioinformatics drive progress in human identification.

Purpose of the Study:

• To develop an efficient deep learning approach for human identification using spatial-temporal gait features.
• To address limitations of current methods in handling gait variations for accurate identification.

Main Methods:

• Utilized a gait energy image (GEI) based Siamese neural network for spatial feature extraction.
• Employed deep 3D convolutional networks (C3D) to learn temporal gait features.
• Integrated spatial and temporal features using Null Foley-Sammon Transform (NFST) and distance metric learning.

Main Results:

• The proposed framework effectively extracts robust and discriminative spatial-temporal gait features.
• NFST and metric learning improved the separation of gait features from the same and different individuals.
• Experiments on a large gait database demonstrated superior performance compared to state-of-the-art methods.

Conclusions:

• The developed deep learning framework offers a significant advancement in gait-based human identification.
• The combination of GEI, C3D, NFST, and metric learning provides a powerful solution for handling gait variations.
• This approach shows high potential for real-world biometric security applications.