Vision-Based Moving Obstacle Detection and Tracking in Paddy Field Using Improved Yolov3 and Deep SORT

Area of Science:

• Agricultural Engineering
• Computer Vision
• Robotics

Background:

• Intelligent agricultural machines require robust obstacle avoidance systems for autonomous operation in complex environments like paddy fields.
• Detecting and tracking moving obstacles is crucial for enhancing the intelligence and safety of agricultural machinery.

Purpose of the Study:

• To develop and evaluate a machine vision system for detecting and tracking moving obstacles in paddy fields.
• To improve the performance and processing speed of obstacle detection and tracking for agricultural robots.

Main Methods:

• A machine vision system was developed using a red, green, and blue (RGB) camera and computer, mounted on a transplanter.
• An improved You Only Look Once version 3 (Yolov3) algorithm combined with deep Simple Online and Realtime Tracking (deep SORT) was utilized for obstacle detection and tracking.
• The improved Yolov3 features architectural modifications, including 23 residual blocks and single upsampling, along with novel loss calculation functions.

Main Results:

• The improved Yolov3 achieved a mean intersection over union (mIoU) score of 0.779 on a custom dataset of moving obstacles (humans, water buffalo).
• The system demonstrated a 27.3% increase in processing speed compared to the standard Yolov3.
• Real-world paddy field tests showed an average processing speed of 5-7 frames per second (FPS), meeting operational demands.

Conclusions:

• The proposed machine vision system effectively detects and tracks moving obstacles in paddy fields, enhancing agricultural machine intelligence.
• The improved Yolov3-deep SORT method offers a significant speed advantage and acceptable detection accuracy for practical applications.
• This system provides a foundation for more flexible autonomous navigation in intelligent agriculture.