Image Feature Detectors in Agricultural Harvesting: An Evaluation.

Area of Science:

• Computer Vision
• Agricultural Technology
• Robotics

Background:

• Image feature detection is crucial for agricultural harvesting, but challenging in harsh environments due to elusive Ground Truth.
• Existing methods struggle to assess feature detectors without reliable Ground Truth data in agricultural settings.

Purpose of the Study:

• To develop and validate a Ground Truth-independent approach for assessing feature detectors in agricultural harvesting.
• To create and release the first agricultural harvesting dataset covering rice, corn, soybean, wheat, and rape.
• To identify the most effective feature detectors for agricultural harvesting applications.

Main Methods:

• Assembled a novel agricultural harvesting dataset with four crop types.
• Developed a Ground Truth-independent assessment methodology focusing on efficiency, repeatability, and feature distribution.
• Evaluated eight distinct feature detectors using the created dataset.

Main Results:

• The FAST detector and ASLFeat demonstrated superior performance in agricultural harvesting scenarios.
• The developed Ground Truth-independent method proved effective for detector evaluation.
• The new dataset provides a valuable resource for agricultural computer vision research.

Conclusions:

• FAST and ASLFeat are recommended for agricultural harvesting applications.
• The Ground Truth-independent assessment approach offers a reliable framework for future detector evaluations.
• This work facilitates improved feature extraction techniques for automated crop reaping.