Optimizing Fruit Yield Prediction: Evaluating Multiobject Tracking Algorithms for Calamansi Fruit Detection Using Yolov8m

Accurate fruit yield prediction is necessary for enhancing agricultural methods and providing quality output. This work analyzes the efficacy of multi-object tracking (MOT) algorithms in detecting calamansi fruits from drone-captured video footage, using the YOLOv8m object detection model. Data were gathered from a commercial calamansi farm in Naujan, Philippines, and prepared for model training. Three MOT algorithms-SORT, DeepSORT, and ByteTrack-were evaluated using performance metrics such as Multiple Object Tracking Accuracy (MOTA), Higher Order Tracking Accuracy (HOTA), and ID F1 Score (IDF1). ByteTrack achieved exceptional performance with MOTA, HOTA, and IDF1 ratings of 59.67 %, $\mathbf{6 5. 4 2 \%}$, and $\mathbf{7 3. 7 2 \%}$ respectively. The integration of YOLOv8m for detection with ByteTrack for tracking significantly reduced double-counting while ensuring good computational efficiency. The results underscore ByteTrack's capability in real-time fruit monitoring and its significance for improving yield estimate and agricultural production.