Abstract:
Deep learning algorithms, particularly Convolutional Neural Networks (CNNs), have
become the state-of-the-art techniques for object detection, classification, segmentation,
and behavior classification. Despite their extensive application across various domains,
including agriculture, cow identification in dairy farming often relies on time-consuming,
costly, and inaccurate methods such as direct visual monitoring or invasive contact
devices like sensors. This research evaluates the performance of different variants of
the YOLOv8 and YOLOv9 models for cow detection and counting in a dairy farming
context. Our dataset, consisting of 2,956 cow images extracted from videos provided by
Nestlé dairy farms, was processed using the Roboflow platform to convert video frames
into images. These images were augmented to simulate varying illumination conditions
and annotated using the RoboFlow tool in YOLO format. Among the YOLOv8 variants,
the YOLOv8l model achieved superior performance, with a mean Average Precision
(mAP) of 91.11%, precision of 92%, and recall of 89%. Additionally, the YOLOv8l
model demonstrated effective cow counting capabilities within individual frames. In
comparison, the YOLOv9l model showed a mAP of 90%, indicating its robustness in cow
detection tasks. Our findings highlight the YOLOv8l model’s potential in enhancing cow
identification and counting processes in dairy farming, offering a reliable, non-invasive,
and efficient alternative to traditional methods. This research underscores the practical
implementation of advanced deep learning models in livestock management, contributing
significantly to the future research directions of the YOLO algorithm.
