Nowadays, the treatment costs associated with lameness rank second among common diseases of cattle. The standard method for detecting lameness is visual observation of the herd by the farmer. However, these methods are time-consuming and labor-intensive and, due to the qualitative nature of the assessment, involve many discrepancies between different human assessors. This study aims to develop fully automated end-to-end methods for the video-based assessment of lameness in dairy cows using data science. For the study, 832 cows with varying degrees of lameness were recorded. The video recordings were then divided into individual frames, where deep learning detected a single cow and its characteristic anatomical points. A custom 7-point locomotion scoring system, inspired by the commonly used 5-level Sprecher (Zinpro) scale, was introduced and evaluated. This scale was used to assess lameness severity based on processed data, which were analyzed using an expert system, machine-learning methods, and a deep-learning approach. Our solution is based on the analysis of the spine curvature, head position, and distance between pairs of legs. The accuracy of detecting binary lameness (healthy vs. lame) through multiple locomotion features approaches expert-level performance, at 0.821 and 0.872, respectively.
Szyc et al. (2025) studied this question.