Research on Multi Perspective Horse Behavior Recognition and Liming Detection Method Based on Deep Learning

In the field of equestrianism, automated analysis of horse behavior is crucial for health risk management and training optimization. This article proposes a deep learning based unlabeled multi view video analysis framework that integrates convolutional neural networks (CNN) and bidirectional long short-term memory networks (BiLSTM) to achieve daily behavior classification and limp detection of horses. By preprocessing 172 multi view videos of 133 horses, extracting coordinates of 12 key parts and normalizing them, a spatiotemporal feature joint modeling architecture was constructed. The experimental results show that the model achieves a 10 fold cross validation accuracy of 93.2% in 15 types of behavior classification, and a leave one method validation accuracy of 86.0%; In limp detection, the accuracy of the multi view fusion scheme reached 87.9%, with an AUC-ROC of 0.913. The ablation experiment showed that the BiLSTM module improved classification accuracy by 4.5% compared to a single CNN, verifying the necessity of temporal modeling. This method provides a low-cost and highly robust automated monitoring solution for equestrian scenes, which can effectively assist in early health anomaly recognition and training effectiveness evaluation.