Field-road classification for agricultural vehicles in China based on pre-trained visual model

Area of Science:

• Agricultural Engineering
• Computer Vision
• Robotics

Background:

• Accurate classification of agricultural vehicle trajectories (field vs. road) is crucial for analyzing operational behavior.
• Existing methods often struggle to capture the nuanced movement patterns specific to agricultural activities.

Purpose of the Study:

• To develop a novel multi-view field-road classification method for Global Navigation Satellite System (GNSS) trajectories.
• To enhance the analysis of agricultural vehicle behavior by accurately distinguishing between in-field and on-road movements.

Main Methods:

• Proposed a multi-view approach extracting physical and visual feature vectors for each trajectory point.
• Utilized a pre-trained ResNet model, fine-tuned with trajectory point images for effective visual feature extraction.
• Generated images from trajectory points and their neighbors to provide contextual information for the model.

Main Results:

• Achieved high accuracy rates: 92.56% (Wheat 2021), 87.91% (Paddy), 90.31% (Wheat 2023), and 94.23% (Wheat 2024).
• Outperformed state-of-the-art methods by 2.99% to 4.42% in F1-score across four datasets.
• Experimental results validated the necessity and effectiveness of the proposed multi-view classification strategy.

Conclusions:

• The multi-view field-road classification method offers a robust and accurate solution for agricultural vehicle behavior analysis.
• The fine-tuned ResNet model effectively leverages both general and task-specific knowledge for improved trajectory classification.
• This approach provides a significant advancement in understanding and analyzing agricultural machinery operations.