From Body Parts Responses to Underwater Human Detection
Thesis Topic
- Motivation: Underwater surveillance system
- Problem: Handle body occlusion issue in underwater enviroment
- Solution: Detect body part instead of whole human body
- Model: A multi-class body part detector for detecting head, arm, torso and leg
Object Detection Model
- faster R-CNN, SSD and YOLOv2
- TensorFlow Object Detection API -> implementation of faster R-CNN and SSD
- darkflow -> implementation of YOLOv2
Underwater Human Body Parts Dataset
- 4151 labelled images (including images and video screenshots)
- Data argumentation (horizontal flip)
- labelImg Tool
Detection Results and Performance
*detection speed depends on the GPU of device
- Faster R-CNN: the highest detection performance, relatively slow speed
- SSD: decent detection performance and detection speed
- YOLOv2: the fastest detection speed but poor detection performance
