Object Detection for Agricultural Vehicles: Ensemble Method Based on Hierarchy of Classes.

Area of Science:

• Computer Vision
• Agricultural Robotics
• Machine Learning

Background:

• Autonomous agricultural vehicles require robust vision-based object detection for safe operation.
• Limited labeled agricultural datasets hinder the direct application of state-of-the-art object detectors.
• Existing models often struggle with distinguishing between similar agricultural object classes.

Purpose of the Study:

• To address the challenge of limited labeled data in agricultural object detection.
• To improve the accuracy and reliability of object detection for autonomous agricultural vehicles.
• To develop an ensemble method for combining multiple object detection models.

Main Methods:

• Utilized two YOLOv5 object detection models: one general-purpose pre-trained, one agriculture-specific.
• Proposed a novel ensemble module employing a hierarchical class structure for combining model detections.
• Evaluated the ensemble's performance on a dedicated agricultural test dataset.

Main Results:

• The proposed ensemble module significantly increased mean Average Precision (mAP@.5) from 0.575 to 0.65.
• The ensemble method effectively reduced misclassifications between similar object classes.
• Translating detections to higher hierarchical levels further boosted mAP@.5 to 0.701, albeit with reduced granularity.

Conclusions:

• The developed ensemble module enhances object detection performance in agriculture.
• Hierarchical class structures offer a promising approach for improving agricultural object detection accuracy.
• The findings contribute to the advancement of autonomous agricultural vehicle technology.