PPE-Detection-Pose

Pose guided anchoring for detecting proper use of personal protective equipment

Introduction
Implementation
CPPE Dataset
Comparative studies: Pre-trained object detection models

Introduction

The repository presents the implementation of vision-based PPE compliances (e.g., hardhats, safety vests) for workers.

^{Overview of the pose-enhanced framework for Non-PPE use recognition}

Implementation

Worker pose estimation

Worker pose estimation mainly follows the work of OpenPose to detect worker body parts as spatial anchors.

Download pose visualization images for CPPE Dataset from Google Drive.
Download pose estimation results in JSON formats from Google Drive.

Part attention localization

The detected keypoints can guide the localization of part attention regions depending on the types of PPE items.

^{Overview of the part attention localization module}

COCO keypoint output format

{0,  "Nose"}, {1,  "Neck"}, {2,  "RShoulder"}, {3,  "RElbow"}, {4,  "RWrist"}, {5,  "LShoulder"}, {6,  "LElbow"}, {7,  "LWrist"}, {8,  "RHip"}, {9, "RKnee"}, {10, "RAnkle"}, {11, "LHip"}, {12, "LKnee"}, {13, "LAnkle"}, {14, "REye"}, {15, "LEye"}, {16, "REar"}, {17, "LEar"}

To navigate through these image patches, body knowledge-based rules are defined using detected 2D keypoints to configure the location and size of the objects’ bounding boxes under various workers’ orientations.

^{Head attention regions for hardhat recognition (left image) and Upper-body attention regions for safety vest recognition (right image)}

Binary classification for PPE and non-PPE use

Generally, each worker produces two types of part attention regions: head attention region and upper-body attention region if their body parts are visible in the image. To recognize PPE instances in the body part attention regions, two classifiers - hardhat classifier f1(X1) and vest classifier f2(X2), are constructed.

Data and weights are available in Google Drive. (hardhat vest)
The following pretrained CNN classifier models are available.
- VGG-16 (hardhat vest)
- VGG-19 (hardhat vest)
- MobileNet (hardhat vest)
- ResNet-18 (hardhat vest)
- ResNet-34 (hardhat vest)
- ResNet-101 (hardhat vest)
- ResNet-152 (hardhat vest)

^{hardhhat recognition(left image) and vest recognition(right image)}

CPPE Dataset

Construction Personal Protective Equipment (CPPE) Dataset
- The proposed CPPE dataset consists of 932 images, including 2,747 instances of hardhats, 1,339 instances of safety vests, and 3,428 workers by collecting data from Pictor-v3 dataset, GDUT-HWD dataset, Safety helmet wearing detect dataset (SHWD), and web-mined images. Data available in Google Drive.
Related public PPE datasets
- The Pictor-v3 dataset contains 774 images for public access. Data available in Google Drive.
- The GDUT-HWD dataset contains 3,174 images, which has been split into 1,587 for training (GDUT-HWD trainval) and 1,587 for testing (GDUT-HWD test). It contains 18,893 instances. Data available in Baidu Yun (pwd:dstk).
- The SHWD contains 7581 images with 9,044 safety helmet wearing objects(positive) and 111,514 normal head objects(not wearing or negative). Data available in Google Drive.

Comparative studies: Pre-trained object detection models

The following pretrained models on the CPPE dataset are available.

The python demo is used for the quick results preview and test.

Citation

If you find this dataset useful in your research, please consider cite:

@article{xiong2021pose,
  title={Pose guided anchoring for detecting proper use of personal protective equipment},
  author={Xiong, Ruoxin and Tang, Pingbo},
  journal={Automation in Construction},
  volume={130},
  pages={103828},
  year={2021},
  doi = {https://doi.org/10.1016/j.autcon.2021.103828}
}