Visible and Thermal Image-Based Trunk Detection with Deep Learning for Forestry Mobile Robotics.

Area of Science:

• Forestry science
• Robotics
• Computer Vision

Background:

• Forest wildfires necessitate in-situ forest inventory and biomass management.
• Mobile robotics offer a potential solution for these challenges.

Purpose of the Study:

• To develop and evaluate deep learning models for ground-level tree trunk detection in forests.
• To create and release a public forestry dataset for model training and benchmarking.

Main Methods:

• A forestry dataset of 2895 images was created.
• Five object detection models (SSD MobileNetV2, SSD Inception-v2, SSD ResNet50, SSDLite MobileDet, YOLOv4 Tiny) were trained and benchmarked.
• Model performance was evaluated using Average Precision (AP) and F1 score, with inference time measured on CPU and GPU.

Main Results:

• YOLOv4 Tiny demonstrated superior performance with the highest AP (90%) and F1 score (89%).
• YOLOv4 Tiny achieved the fastest inference time at 8 ms on GPU.
• All benchmarked models showed promising results for tree trunk detection.

Conclusions:

• Deep learning object detection is effective for identifying tree trunks in forest environments.
• YOLOv4 Tiny is a highly efficient model for real-time tree trunk detection in forestry applications.
• This research advances vision perception systems for autonomous forestry robots.