AI_contest

• Monitoring Safety Equipment through a object detection model
• 객체 탐지 모델을 통한 공사 현장 안전 장구류 착용 모니터링 서비스

Contents

1. Tool / Dataset
2. Training
3. demo

1. Tools / Dataset

Dataset

• AI Hub (공사현장 안전 장비 인식 이미지)
  • https://aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&aihubDataSe=realm&dataSetSn=163
• total 6000 images
• classes to detect(6 classes)
  • 안전 벨트 착용 여부(착용/미착용)
  • 안전화 착용 여부(착용/미착용)
  • 안전모 착용 여부(착용/미착용)

Tools

2. Training

Model - YOLO for read-time detection

Problem

1. Class Imbalance

   • Solution

     → Data Augmentation (객체 기반 Crop, Rotation(180°))

     ※ Reference

     

2. Especially hard to detect "Hard / No hard , Belt / No Belt" classes

   • hard to detect Hard/No Hard classes(안전화 착용 여부 탐지) since safety shoes are ..
     1. small objects
     2. A small difference between safety shoes and sneakers
   
   

   Solution

   → Not just resize Cropped images into 640 x 640. Apply Super Resolution techniques on cropped images. Our team thought that restoring cropped images into high resolution images by super resolution techniques can lead to a high-performance.

   → SRCNN(Super Resolution Convolutional Neural Network)

Result

• 전반적인 mAP 결과

• 클래스 별 결과

3. Demo(prototype)

framework
Streamlit - development of a prototype
library
openCV - input real-time images from the webcam into the model