System of Counting Green Oranges Directly from Trees Using Artificial Intelligence

Abstract

Agriculture is one of the most essential activities for humanity. Systems capable of automatically harvesting a crop using robots or performing a reasonable production estimate can reduce costs and increase production efficiency. With the advancement of computer vision, image processing methods are becoming increasingly viable in solving agricultural problems. Thus, this work aims to count green oranges directly from trees through video footage filmed in line along a row of orange trees on a plantation. For the video image processing flow, a solution was proposed integrating the YOLOv4 network with object-tracking algorithms. In order to compare the performance of the counting algorithm using the YOLOv4 network, an optimal object detector was simulated in which frame-by-frame corrected detections were used in which all oranges in all video frames were detected, and there were no erroneous detections. Being the scientific and technological innovation the possibility of distinguishing the green color of the fruits from the green color of the leaves. The use of YOLOv4 together with object detectors managed to reduce the number of double counting errors and obtained a count close to the actual number of oranges visible in the video. The results were promising, with an mAP50 of 80.16%, mAP50:95 of 53.83%, precision of 0.92, recall of 0.93, F1-score of 0.93, and average IoU of 82.08%. Additionally, the counting algorithm successfully identified and counted 204 oranges, closely approaching the actual count of 208. The study also resulted in a database with an amount of 644 images containing 43,109 orange annotations that can be used in future works.

Keywords

• fruits account
• deep learning
• YOLO
• object tracker