Accurate fruit counting is a key component of yield estimation in precision agriculture, enabling the optimization of harvesting, logistics, and export processes. In this work, we propose a system for the automatic counting of avocados in videos captured in real agricultural fields, combining YOLOv8m for fruit detection with the BoT-SORT algorithm for temporal tracking. The system was validated using a custom dataset composed of public images and original footage collected in commercial Peruvian orchards, under real-world conditions of lighting, occlusion, and high fruit density. The detection model achieved a mAP@0.5 of 0.953, a recall of 0.888, and an F1-score of 0.919. For cumulative counting, the system reached an Absolute Counting Precision (ACP) of 92.68%, along with a Mean Absolute Error (MAE) of 2.0 and a Root Mean Square Error (RMSE) of 3.16, reducing duplicate counts through association and re-identification adjustments. These results approach the state of the art in controlled environments, such as the work of Du et al.  (2024), who reported an ACP of 95.33% and an MAE of 3.33. The proximity of our results, despite the more challenging capture conditions, demonstrates that a design focused on cumulative counting enables robust metrics and outcomes applicable to real-world agricultural scenarios, offering a practical solution for producers and companies in the sector.