Automated Region of Interest-Based Data Augmentation for Fallen Person Detection in Off-Road Autonomous Agricultural...

Area of Science:

• Agricultural Automation
• Computer Vision
• Robotics

Background:

• Growing agricultural demand necessitates autonomous vehicles.
• Fallen person detection is crucial for safety in autonomous agricultural operations.
• Limited off-road datasets hinder the performance of detection models.

Purpose of the Study:

• To address data scarcity for fallen person detection in off-road agricultural environments.
• To propose and evaluate an automated data augmentation technique for improving detection generalization.
• To enhance the safety and reliability of autonomous agricultural vehicles.

Main Methods:

• Developed the Automated Region of Interest Copy-Paste (ARCP) technique.
• Utilized YOLOv8x-seg and Grounded-Segment-Anything for segmentation annotation generation.
• Applied ARCP to copy-paste fallen person objects onto off-road backgrounds, creating augmented datasets.

Main Results:

• ARCP significantly improved detection accuracy for YOLOv7x (17.8% increase) and YOLOv8x (12.4% increase).
• Achieved high detection accuracy rates of 95.6% and 96.2% respectively.
• Demonstrated the effectiveness of ARCP in overcoming data limitations in off-road scenarios.

Conclusions:

• The ARCP technique is highly applicable for enhancing object detection in data-scarce off-road environments.
• This advancement is expected to significantly impact object detection technology in the agricultural industry.
• Improved fallen person detection contributes to safer autonomous agricultural operations.