Automated Fruit Counting with YOLOv5 Model for Harvest Management in Orchards

Abstract

Proper counting of fruits on trees is vital for proper timing & scheduling of harvesting, labor force requirement and making agricultural sector more efficient. Manual methods are costly and time-consuming together with being inaccurate in most if not all aspects hence the need for automation. It also demonstrates how real-time fruit detection and counting can be performed using YOLOv5 with potential in handling orchards in a more efficient manner. YOLOv5 was trained on a dataset of fruits in different settings including clear sun lit conditions, in overcast conditions, and in conditions where fruits are entangled in large vegetation. The proposed model was able to release the highest detection accuracy with higher precision, recall, and F1 scores; 97.5%, 96.1%, respectively and 96.8% for the F1-score. It also attained a mean Average Precision ([email protected]) of 98.2%, and further validating its endurance in fruit detection regardless of scenarios. Due to the YOLOv5's real-time indexing that provides about 30 FPS on edge devices, that makes it ideal to be used on drones and mobile autonomous vehicles in large-scale orchards. A comparison with other object detection models like Faster R-CNN, SSD, and Mask R-CNN, proved that YOLOv5 performs better in accuracy as well as in time efficiency. However, there was a small decline in the detection accuracy in cases of partial fruit occlusion or dense foliage. In conclusion, this particular work demonstrates that indeed YOLOv5 can easily solve the problem of automated fruit counting and bring multiple enhancements to the orchard management and precision agriculture.

Topics & Concepts