Multiple Object Tracking in Native Bee Hives: A Case Study with Jataí in the Field.

Artificial intelligence approaches, such as computer vision, can help better understand the behavior of bees and management. However, the accurate detection and tracking of bee species in the field remain challenging for traditional methods. In this study, we compared YOLOv7 and YOLOv8, two state-of-the-art object detection models, aiming to detect and classify Jataí Brazilian native bees using a custom dataset. Also, we integrated two tracking algorithms (Tracking based on Euclidean distance and ByteTrack) with YOLOv8, yielding a mean average precision (mAP 50 ) of 0.969 and mAP 50–95 of 0.682. Additionally, we introduced an optical flow algorithm to monitor beehive entries and exits. We evaluated our approach by comparing it to human performance benchmarks for the same task with and without the aid of technology. Our findings highlight occlusions and outliers (anomalies) as the primary sources of errors in the system. We must consider a coupling of both systems in practical applications because ByteTrack counts bees with an average relative error of 11%, EuclidianTrack monitors incoming bees with 9% (21% if there are outliers), both monitor bees that leave, ByteTrack with 18% if there are outliers, and EuclidianTrack with 33% otherwise. In this way, it is possible to reduce errors of human origin.