exploring key visual features for early lameness detection: toward transparent intelligence

Lameness in cattle, characterized by abnormal stride and gait, poses significant economic and welfare challenges in agriculture. traditional visual inspections lack accuracy and scalability, prompting the development of transparent computer vision-based detection systems. this study leverages a dataset of 170 cattle videos from public sources and the university of tehran’s cattle farm, preprocessed into 1226 one-second sub-clips (416×416 pixels, 25 fps) to mitigate noise from unpredictable cattle behavior. using the yolov7 model, we extracted 35 temporal features, including step sizes, speed, acceleration, and relative head-to-leg coordinates, focusing on the cattle’s head, legs, and back. these features were further engineered using time-series characterization techniques and hypothesis testing, yielding 3773 features. a deep learning model, trained on these features, achieved 88.66% accuracy and 93.74% auc, while a light gradient boosting machine model on engineered features reached 81.3% accuracy and 90.8% auc. sensitivity analysis highlighted leg and head-related features as critical for lameness detection. by emphasizing interpretable features and robust modeling, this approach enhances transparency, improving animal welfare and farm productivity under diverse conditions.