Real-time keypoints detection for autonomous recovery of the unmanned ground vehicle

This repo contains training and testing code for our paper on UGV-KPNet, a task for detecting the keypoints of the small UGV form the images captured by a monocular camera to obtain positioning and attitude information of the UGV. More information about the project can be found in our paper.

If you find UGV-KPNet useful in your research, please cite:

@article{ugvkpnet,
  author    = {Li, Jie and Zhang, Sheng and Han, Kai and Yuan, Xia and Zhao, Chunxia and Liu, Yu},
  title     = {Real-time keypoints detection for autonomous recovery of the unmanned ground vehicle},
  journal   = {IET Image Processing},
  volume    = {14},
  number    = {17},
  pages     = {4690--4700},
  year      = {2021},
  publisher = {IET}
}

Dataset

Training

python train_ShuffleNetV2_robot.py \
--json_path='/path/to/json/robot_keypoints_20190426_640x360.json' \
--saved_model='robotkp_LWShuffleNetV2_v1_16_cat_w1.0_v0.1.pth' \
--model='LWShuffleNetV2_v1_16_cat' \
--multistage 0 \
--gpu_ids 3 \
--epochs=400 \
--batch_size=24 \
--workers=4 \
--lr=0.5 \
--batch_size_val=4 \
--workers_val=1 \
--data_augment True \
--resample True \
--logdir='./logs/robotkp_LWShuffleNetV2_v1_16_cat_w1.0' 2>&1 |tee train_LWShuffleNetV2_v1_16_cat_w1.0_v0.1.log

Our work is built on https://github.com/tensorboy/pytorch_Realtime_Multi-Person_Pose_Estimation.