A methodology for detection and localization of fruits in apples orchards from aerial images

TL;DR

This paper introduces a methodology using aerial images and CNNs for accurate apple fruit detection, counting, and 3D localization, improving yield estimation by integrating multi-view geometry.

Contribution

It presents a novel approach combining CNN-based detection with multi-view geometry for fruit localization and counting from aerial images.

Findings

High correlation (>0.8) between estimated fruit count and true yield.

Provides a publicly available annotated dataset for CNN training.

Demonstrates effective fruit tracking and localization in 3D space.

Abstract

Computer vision methods based on convolutional neural networks (CNNs) have presented promising results on image-based fruit detection at ground-level for different crops. However, the integration of the detections found in different images, allowing accurate fruit counting and yield prediction, have received less attention. This work presents a methodology for automated fruit counting employing aerial-images. It includes algorithms based on multiple view geometry to perform fruits tracking, not just avoiding double counting but also locating the fruits in the 3-D space. Preliminary assessments show correlations above 0.8 between fruit counting and true yield for apples. The annotated dataset employed on CNN training is publicly available.