This repository is the official documentation & implementation of VFP290K: A Large-Scale Benchmark Dataset for Vision-based Fallen Person Detection.

Our pretrained models except YOLO are based on MMdetection2 detection framework. You can donwload coco-pretrained models for the transfer learning.

Download our VFP290K dataset in here: VFP290K.

CUDA=11.1
python=3.7
CUDNN=7.6.5

We use pytorch=1.8.0 from this link.

pip install openmim
mim install mmdet
pip install future tensorboard
pip install -r requirements.txt

python make_label.py --data_root_dir <VFP directory>

You can find checkpoints files in the official repository and put them into ./checkpoints

We prepare all config files used in our experiments in "configs/VFP290K".
Set your labels.txt and VFP290K data directory.
(classes= "<YOUR labels.txt DIRECTORY>", data_root = "<YOUR DATA DIRECTORY>")

To train and evaluate the model(s) in the paper, run this command:

• single gpu training

  python tools/train.py <config> --gpu-ids <device> 
  

  and indicate path of the config file and gpu id, respectively. This example is for train faster R-CNN model on gpu 0.
  ex) python tools/train.py configs/VFP290K/faster_rcnn_r50_1x_benchmark.py --gpu-ids 0

• multi gpu training

  bash ./tools/dist_train.sh <config> <num_gpu> 
  

  <num_gpu> is a number of gpus to use. This example is for train faster R-CNN model with 4 gpus.
  ex) bash ./tools/dist_train.sh configs/VFP290K/faster_rcnn_r50_1x_benchmark.py 4

• test After train the model, you can evaluate the result.

  python tools/test.py <config> <weight> --eval bbox --gpu-ids <device>
  

  is the path of the trained model weight.
  ex) python tools/test.py configs/VFP290K/faster_rcnn_r50_1x_benchmark.py work_dirs/faster_rcnn_r50_1x_benchmark/latest.pth --eval bbox --gpu-ids 1

cd yolov5
python configs/data_refactoring.py --data_root_dir <{YOUR DATA ROOT DIRECTORY}/yolov5>

E.g) configs/benchmark.yaml
train: /media/data1/VFP290K/VFP290K/yolov5/benchmark/train/image
val: /media/data1/VFP290K/VFP290K/yolov5/benchmark/val/image

Training process is exactly same with official code. E.g)

python train.py --img-size 640 --epochs 100 --data configs/benchmark.yaml --batch-size 48 --cfg ./models/yolov5x.yaml --device 0,1 --workers 8 

You can train your model by using this script.

Along with the training, test process is also needed to evaluate our model.

python test.py --weights runs/train/exp<your_exp_num>/weights/best.pt --data data/test.yaml --batch-size 48 --img-size 640 --conf-thres 0.5 --iou-thres 0.5 --device 0,1

Our model achieves the following performance on benchmark:

Our model achieves the following performance on Background:

Our model achieves the following performance on light conditions and camera heights:

This repository is copyrighted under GPLv3 license