Abstract
In this study, we developed a dataset of behaviors associated with lameness in dairy cows. The data collection utilized IoT collars that were placed around the necks of 10 dairy cows. This publicly available dataset contains 441 labeled behaviors, amounting to over 7 h of recording time. It includes acceleration data referenced in both body and world frames, as well as gyroscope signals, which facilitated the extraction of 112 relevant features for classifying key behaviors such as walking, grazing, and resting through machine learning algorithms. To enhance model performance and reduce feature dimensionality, automatic feature selection techniques were applied before classification. The dataset's effectiveness was assessed using various classification models, including Support Vector Machines (SVM), Logistic Regression, Decision Trees, and Random Forests. Results indicated that signals referenced to the body frame yielded better behavior discrimination, achieving a maximum macro F1-score of 0.9625 with the SVM model. This public dataset can facilitate early lameness detection by enabling accurate classification of behavior patterns.