[TEAM] Vision AI를 활용한 도로 위 이상 상황 탐지 서비스 개발

■ Description

- 종합 기능형 교통 관제 시스템
- YOLO를 활용하여 Object Detection을 진행하며, 이를 바탕으로 실시간 도로 상황을 인식하고 판단하는 서비스 개발
- 판단 데이터는 종합 기능형 교통 관제 시스템으로 제공

 

■ Used Software

- YOLOv5
- OpenCV
- Pytorch
- ngrok
- FastAPI
- Docker
- MySQL

 

■ Language

- Python
- HTML/CSS

 

■ My Work

- YOLO 모델 학습
- YOLO 객체 인식
- OpenCV 카메라 연결
- 도로영역 탐지
- API 연결

 

■ Service Framework

■ YOLO Object Detection

- Event Type

EVENT/ : YOLO Object Detection을 통해 인지되는 Object를 기반으로 Event_type을 지정
│
├── Human/
│   ├── P01 : person detected
│   └── ...
│
├── Bicycle or Motorcycle/
│   ├── B01 : bicycle detected with person
│   ├── B02 : motorcycle detected with person
│   └── ...
│
├── Fire/
│   ├── F01 : fire detected
│   ├── F02 : smoke detected
│   ├── F03 : car fire detected
│   └── ...
│
├── Animal/
│   ├── A01 : cat or dog detected
│   ├── A02 : deer or elk detected
│   ├── A03 : racoon detected
│   └── A04 : wild boar or pig detected
│
├── Obstacle/
│   ├── S01 : Tree detected on the road
│   ├── S02 : Rock detected on the road
│   ├── S03 : Box detected on the road
│   ├── S04 : Tire detected on the road
│   ├── S05 : Drum detected on the road
│   └── ...
│
└── Vehicle/
    └── ...

 

- Used Dataset

YOLOv5m: https://github.com/ultralytics/yolov5

GitHub - ultralytics/yolov5: YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite

🔥 Fire Detection: https://universe.roboflow.com/porvip/fire-szxx8

fire Object Detection Dataset by PORVIP

🐶 Animal Detection
- Cat: https://universe.roboflow.com/yolo-4akh3/cat-model-oyjob

cat model Object Detection Dataset and Pre-Trained Model by yolo

- Deer: https://universe.roboflow.com/jwellstx-ahvez/deerface

DeerFace Object Detection Dataset and Pre-Trained Model by jwellstx

- Racoon: https://universe.roboflow.com/objectdection2/racoonsfinder

RacoonsFinder Object Detection Dataset and Pre-Trained Model by ObjectDection2

- Dog: https://universe.roboflow.com/tansam-uunrl/dog-detector1

Dog Detector1 Object Detection Dataset and Pre-Trained Model by Tansam

- Wild boar: https://universe.roboflow.com/workspace1-lhfkr/wildboar-afbbo

WILDBOAR Object Detection Dataset by workspace1

📦 Obstacle Detection
- Tree: https://universe.roboflow.com/tree-detection-h9dcy/tree-detection-ekaot

Tree Detection Object Detection Dataset and Pre-Trained Model by Tree Detection

- Rock: https://universe.roboflow.com/aditya-xxyrw/rocks-mrlaw

rocks Object Detection Dataset and Pre-Trained Model by Aditya

- Box: https://universe.roboflow.com/objectdetection-rclje/parceldetection-cdilp

ParcelDetection Object Detection Dataset and Pre-Trained Model by ObjectDetection

- Tire: https://universe.roboflow.com/robert-almalak/tires-9zgkh

Tires Object Detection Dataset and Pre-Trained Model by Robert AlMalak

 

YOLO Framework

 

 

Road Boundary Check

1. 영상 촬영 시작 후 일정 시간 동안 차량 객체가 인식되는 좌표 정보 저장
2. 저장된 좌표 데이터를 Convexhull Algorithm을 활용하여 꼭지점 좌표만 저장
3. Polygon 타입 데이터를 활용하여 도로 영역 선언
4. 영역 탐지 이후 도로 영역에 진입하는 객체에 상황 판단 로직 작동

 

Object Detection

- YOLOv5m (Pretrained model)

→ Total Class: 80

→ Used Classes: Person, Bicycle, Motorcycle, Car, Bus Truck

 

🐶 Animal Detection Model (Custom model)

→ Total Class: 5

→ Used Class: Cat, Dog, Deer/Elk, Racoon, Wild boar/Pig

 

🔥 Fire Detection Model (Custom Model)

→ Total Class: 3

→ Used Class: Fire, Smoke, Car fire

 

📦 Obstacle Detection Model (Custom model)

→ Total Class: 5

→ Used Class: Tree, Box, Tire, Drum can, Rock

 

■ GitHub

GitHub - SCUTUM98/KFQ_TEAM01