Performance Evaluation of Detection and Tracking Algorithms for Automated Grape Cluster Counting

Kemper R.J.H.; Manrique P.A.Z.; Rosas D.D.H.; Gardini S.R.P. · C3 2025 - IEEE Colombian Caribbean Conference (2025) · DOI: 10.1109/c366505.2025.11340275

Resumen

The accurate estimation of crop yield is crucial in modern agriculture; however, traditional methods, such as manual counting of grape clusters, often result in errors exceeding 20% and are impractical for large-scale monitoring under variable environmental and climatic conditions. While object detection techniques have been successfully implemented in realworld agricultural scenarios, the integration and performance of multi-object tracking (MOT) algorithms in such dynamic and unstructured environments remains underexplored. This article provides a comparative evaluation of state-of-the-art object detection and tracking methodologies for use in vineyard conditions. Here, two detection methods-YOLOv12 and RF-DETR-were compared in combination with two tracking algorithms-BoTSORT and ByteTrack-to generate four distinct configurations aimed at the automatic counting of grape clusters. Experimental results on a video dataset recorded in agroindustrial vineyards showed that the combination of YOLOv12 with BoT-SORT achieved the highest average relative counting accuracy, with 84.31%. These findings confirm the potential of AI-based techniques for precision agriculture, offering greater accuracy in yield estimation, optimized resource management, and a substantial reduction in reliance on manual labor.

Instituciones: Universidad Privada Antenor Orrego