Cross-View Gait Recognition Based on U-Net

Updated: April 27th, 2021

Description | Our approach | Results | Getting Started | Citing

Example of invariant Gait Energy Image (GEI) generation.

# Description Inspired by the great successes of GANs in image translation tasks, we propose a new gait recognition technique by using a conditional generative model to generate view-invariant features. The proposed method is evaluated on one of the largest datasets available under the variations of view, clothing and carrying conditions: CASIA gait database B. Experimental results show that the proposed method achieves an outstanding correct classification rate and outperformed state-of-the-art methods specially in carrying-bag and wearing-coat sequences.

In this work, we propose a method of gait recognition using a conditional generative model to generate view-invariant features and overcome appearance variations due to changes of clothing, carrying conditions, and view angle.

Our approach

Framework

Example of invariant Gait Energy Image (GEI) generation.

Conditional GAN (CGAN)

Generator	Discriminator

Results

Next, some qualitative results are shown:

Original GEI representations for the subject 120 on the CASIA-B dataset. Notice that we are showing all the view-angles for three sequences: nm-01, bg-01, and cl-01.

Generated GEI representations for the same subject

Generated GEI representations from multiple subjects on each training step

However, it is necessary to compare our results with other state-of-the-art works.

Comparison with other approaches based on the correct classification rate (CCR)

Getting Started

You can find the notebook here or open it at

Prerequisites

- Tensorflow 2.x
- Keras
- OpenCV
- Numpy
- Matplotlib

Built With

• CASIA - The dataset used
• Pix2Pix - Based on
• GoogleColab - The virtual machine used in the experiments

Citing

Please, cite this work as follows:

I. R. Tiñini Alvarez and G. Sahonero-Alvarez, "Cross-View Gait Recognition Based on U-Net," 2020 International Joint Conference on Neural Networks (IJCNN), Glasgow, United Kingdom, 2020, pp. 1-7, doi: 10.1109/IJCNN48605.2020.9207501.

Acknowledgments

This work has been inspired on:

• Image-to-Image Translation with Conditional Adversarial Networks
• U-Net: Convolutional Networks for Biomedical Image Segmentation

The following articles have been used to compare our approach

• GaitGANv2: Invariant Gait Feature Extraction Using Generative Adversarial Networks
• GaitGAN: Invariant Gait Feature Extraction Using Generative Adversarial Network
• View invariant gait recognition using only one uniform model

Authors

• Israel Tiñini Alvarez
• Guillermo Sahonero-Alvarez