Construction-Site-Safety-PPE-Detection

PPE Detection for Construction Site Safety using YoloV8

4,764 workers died on the job in 2020 (3.4 per 100,000 full-time equivalent workers). Workers in transportation and material moving occupations and construction and extraction occupations accounted for nearly half of all fatal occupational injuries (47.4 percent), representing 1,282 and 976 workplace deaths, respectively.

Occupational Safety and Health Administration (US Department of Labour)

Introduction

There have been various accidents in construction sites, due to the lack of safety equipments for workers. The aim of this project was to detect PPE on a worker, which can be further used for tracking and triggerring alarm (safety monitoring) in future. We use the dataset provided by Roboflow on Construction Site Safety Image Dataset.

For easier use the dataset is already uploaded here: Kaggle Dataset.

The dataset consists of 2801 image samples with labels in YoloV8 format. These images are split into train: 2605, valid: 114 and test: 82 sets. Each folder consists of images and labels folders.

There are 10 classes to detect from the dataset:

‘Hardhat’, ‘Mask’, ‘NO-Hardhat’, ‘NO-Mask’, ‘NO-Safety Vest’, ‘Person’, ‘Safety Cone’, ‘Safety Vest’, ‘machinery’, ‘vehicle’

Setup

The code was run on Kaggle, with a P100 GPU. We installed the ultralytics library by Ultralytics to run YoloV8 custom object detection on the dataset.

For more information check out this notebook: Check yo’self before you wreck yo’self - CSS EDA.

Note: This repository contains all the results, visualizations and best model after custom training on the dataset.

File Hierarchy

1. data folder consists of the yaml file required for training. It also contains 3 folders train, valid and test. Each of these folders have 2 subfolders images (with .jpg files) and labels (with .txt annotations).
2. results folder consists of the prediction results of the model, confusion matrix plot, visualizations of the train and valid batches and PR curves.
3. models folder consists of 2 models, yolov8n.pt which is the pre-trained model on COCO128.yaml and best.pt which is the custom trained yolov8n model on our dataset.
4. source_files folder consists of videos and images for evaluation of our custom trained model.
5. output folder consists of output produced by our custom object detection model after 100 epochs of training.

Code

├───.ipynb_checkpoints
├───assets
├───data
├───├──data.yaml
├───├──ppe_data.yaml
│   ├───test
│   │   ├───images
│   │   └───labels
│   ├───train
│   │   ├───images
│   │   └───labels
│   └───valid
│       ├───images
│       └───labels
├───models
├───output
│   └───output_yolov8n_100e
├───results
└───source_files

Results

The training of YoloV8n model was done for 100 epochs and was completed in 2.719 hours. After training, we get the following results:

Outputs

Future Work

1. Train the model for more epochs.
2. Compare with 4 other models by YoloV8.
3. Create ID tracking of workers and save bounding boxes of workers not wearing proper PPE.
4. ML App deployment with alarm trigerring.