Localization of grazing cows using Yolov8 object detection and known landmarks

Ensuring adequate feed intake is a major challenge for farmers when adopting grazing in the management of dairy cows, despite the recognized welfare benefits of pasture access. This study aims to develop a vision-based system for localizing grazing cows in pastures as the rise of AI and edge computing offers great capabilities to assist in feed management. A total of 3039 labeled images of cows on the pasture were used to train the YoloV8 model, with 62% for training, 26% for validation, and 12% for testing. The dataset was augmented by applying geometric transformations and variations in lighting and contrast before the model was trained to achieve a model that is robust to varying light conditions and positions of the cow in the image. Training Yolov8 nano, small, and medium models on this dataset yielded promising results. mAP50 values were consistently high (>0.87) across training, validation, and test sets for all model sizes. The nano version seems the best suited for applications in a pasture due to less computation power requirements than the other models. The image-based localization achieved an accuracy of 0.89 ± 0.50m for cows in a pasture. A sensitivity analysis of the localization algorithm also indicated that camera-based localization of the cow in a pasture is sensitive to changes in the height of the detected cow, but advances in animal pose estimation might benefit the investigated localization approach. Future research should focus on independent datasets to validate the models and quantify overfitting and on incorporating behavior identification to achieve an integrated monitoring system.