Abstract
Top-view systems for lameness detection have advantages such as easy installation and minimal impact on farm work. However, the unclear lameness motion characteristics of the back result in lower recognition accuracy for these systems. Therefore, we analysed the compensatory behaviour of cows based on top-view walking videos, extracted compensatory motion features (CMFs), and constructed a model for recognising lameness in cows. By locating the hook, pin, sacrum, and spine positions, the motion trajectories of key points on the back were plotted. Based on motion trajectory analysis of 655 samples (258 sound, 267 mild lameness, and 130 severe lameness), the stability mechanisms of back movement posture were investigated, compensatory behaviours in lame cows were revealed, and methods for extracting CMFs were established, including swing and posture features. The feature correlation among differently scoring samples indicated that early-stage lame cows primarily exhibited compensatory swing, while those with severe lameness showed both compensatory swing and posture. Lameness classification models were constructed using machine learning and threshold discrimination methods, achieving classification accuracies of 81.6% and 83.05%, respectively. The threshold method reached a recall rate of 93.02% for sound cows. The proposed CMFs from back depth images are highly correlated with early lameness, improving the accuracy of top-view lameness detection systems.