Detecting lameness in dairy cattle is essential to mitigating the effects of a significant animal welfare and health issue for the dairy industry across diverse farming systems. The locomotion score, a standard method for lameness evaluation, depends on visual assessments conducted by experienced and skilled observers, which limits the objectivity of diagnoses on large-scale farms. In this paper, we propose a gait anomaly detection method through motion reconstruction, which is based on identifying low-dimensional subspaces derived from normal motion patterns within the training data. Specifically, we first generate smooth motions from the trajectories of each key point extracted by a skeleton extraction network in a video. We then use a self-expressive model to learn subspaces from normal motions and reconstruct the given test motion. Our reconstruction module leverages the insight that as the subspace-based approximation strategy only enables reproducing the normal motions, the anomalous motions would induce a significant reconstruction error. Experimental results using cattle gait dataset demonstrate the effectiveness of the proposed method through quantitative and qualitative evaluation.
