2VGait

• This is the official implementation of our paper, Learning Viewpoint-Invariant Features for LiDAR-Based Gait Recognition.
• 2VGait is an end-to-end framework for gait recognition using 3D LiDAR, robust to variations in viewing angles and measurement distances.
• You can also implemenet the previous version, 2V-Gait: Gait Recognition using 3D LiDAR Robust to Changes in Walking Direction and Measurement Distance.

Overview

Overview of the 2VGait, which learns two viewpoint-invariant gait shapes in varying point cloud densities using an attention-based approach.

Data collection

Data acquisition environment with two distances measured from a VLP-32C, which is visualized in a 3D point cloud format.

Experimental result

Comparison with prior studies for evaluating by limiting viewing angles (%).

Usage

Prepare the dataset

Our dataset, KUGait30 is not publicly available at the moment. However, you can refer to our implementation and adapt the dataset generation process for your own data using this codebase.

Training

Convert the 3D pedestrian point clouds into depth-based representation and build the training sets:

python build_dataset_for_training.py \
  --dataset_path ./datasets/KUGait_VLP32C_2022-Spring-C/ \
  --yml_path ./configs/KUGait_VLP32C_2022-Spring-C/build_datasets_1020m.yml

Train a model:

python train.py \
  --dataset_path ./datasets/KUGait_VLP32C_2022-Spring-C/train \
  --model_path ./pretrained_20221201 \
  --batchsize 42 \
  --nepoch 50

Evaluation

Build the test sets:

python build_dataset_for_test.py \
  --dataset_path ./datasets/KUGait_VLP32C_2022-Spring-C/ \
  --yml_path ./configs/KUGait_VLP32C_2022-Spring-C/build_datasets_1020m.yml

Evaluate the trained model:

python test.py \
  --model_path ./pretrained_20221201/model_state_dict.pth \
  --test_path ./datasets/KUGait_VLP32C_2022-Spring-C/test

Citation

If you find this repository useful for your work, please cite our paper:

@article{access2023_ahn,
  title   = {Learning Viewpoint-Invariant Features for LiDAR-Based Gait Recognition},
  author  = {Ahn, Jeongho and Nakashima, Kazuto and Yoshino, Koki and Iwashita, Yumi and Kurazume, Ryo},
  journal = {IEEE Access},
  volume  = {11},
  number  = {},
  pages   = {129749-129762},
  year    = {2023},
  paper   = {https://doi.org/10.1109/ACCESS.2023.3333037}
}

@inproceedings{sii2022_ahn,
  title     = {2V-Gait: Gait Recognition using 3D LiDAR Robust to Changes in Walking Direction and Measurement Distance},
  author    = {Ahn, Jeongho and Nakashima, Kazuto and Yoshino, Koki and Iwashita, Yumi and Kurazume, Ryo},
  booktitle = {Proceedings of the IEEE/SICE International Symposium on System Integration (SII)},
  year      = {2022},
  pages     = {602--607},
  paper     = {https://doi.org/10.1109/SII52469.2022.9708899}
}