CS406 PPE Violation Detection

• Professor: ThS. Cáp Phạm Đình Thăng ([email protected])
• Course Id: CS406.P11 (Xử lý ảnh và ứng dụng)

Team Members

Name	MSSV	Roles
Nguyễn Hữu Nam	22520917	Leader
Nguyễn Trần Phúc	22521135	Member
Hồ Trọng Duy Quang	22521200	Member

A glimpse of the project

Figure 1: Overview of how Violation Detection works on image

Figure 2: Overview of how Violation Detection works on video

Figure 3: Project pipeline

Figure 4: Web demo

Overview: this project focus on detecting PPE (Hardhat, Helmet, Gloves) and classify each person if they are missing one of these

• Image: detect + classify violation
• Video: detect + track + classify violation (video result: drive)

Project structure

CS406-PPE-detection/
├── data/
│   ├── data-ppe.yaml
│   └── split/      #contain train, val, test
├── logs/
├── notebooks/
├── output/
├── sample/
├── scripts/
│   ├── detect_faster_rcnn.py
│   ├── detect_yolo.py
│   ├── loader_faster_rcnn.py
│   └── tracker_yolo.py
├── src/
│   ├── loader/
│   ├── models/
│   ├── parsers/
│   ├── trackers/
│   └── utils/
├── tools/
├── web/
│   ├── app.py
│   └── output/     #output of web
├── weights/
│   ├── best_faster_rcnn.pt
│   └── best_yolo.pt
├── README.md
├── requirements.txt
└── setup.py

I. How to build

1. Installing

git clone https://github.com/Beeditor04/CS406-PPE-detection.git
cd CS406-PPE-detection
pip install -r requirements.txt
python setup.py develop

2. Preparing model

you can download our pretrain faster-rcnn and yolov5n model here:

• faster-rcnn: drive
• yolov5n: drive

And then put it in folder weights.

3. Dataset

• SH17 dataset: kaggle
• Our dataset (put it in data/): kaggle
• data-ppe.yaml: drive

II. Deploy web

streamlit run web/app.py

III. Model

• remember to create .yaml file, put it in the folder data/ or wherever you want

train Faster RCNN

python scripts/train_faster_rcnn.py --data_dir "data/split" --batch_size 8 --epochs 10 --eval_every 5 --iter_every 5 --num_classes 7 --yaml "data/data-ppe.yaml" --lr 0.005 --resize 640 --is_aug 0

train YOLOv5n

• (it've done in jupyter notebook, in notebooks/train_yolo)

run 1 image Faster RCNN inference

python scripts/detect_faster_rcnn.py --weights weights/best_faster_rcnn.pt --img_path sample/images/1.jpg

run 1 image YOLOv5n inference

python scripts/detect_yolo.py --weights weights/best_yolo.pt --img_path sample/images/1.jpg

run tracker YOLOv5n on video

python scripts/tracker_yolo.py --weights weights/best_yolo.pt --vid_dir sample/videos/1.mp4

run tracker Faster RCNN on video

python scripts/tracker_faster_rcnn.py --weights weights/best_faster_rcnn.pt --vid_dir sample/videos/1.mp4

---

IV. Dataset tools (tools/)

• split dataset

python scripts/split_dataset.py --data_dir data/ --output_dir data/split/ --train_size 0.7 --test_size 0.1

• split dataset 10% for exps

python scripts/split_dataset_01.py --data_dir data/ --output_dir data/split_01/ --train_size 0.7 --test_size 0.1

• count dataset

python scripts/count.py --main data/split

• valid dataset

python scripts/is_valid_dataset.py