Computer vision applied to behavior identification for horses

Introduction: The evaluation by video monitoring allows the identification of horse behaviors in a non-invasive and low-cost way.

Materials and Methods: Horses were housed in stalls with high image resolution cameras with infrared capabilities. Recordings were conducted in 5 consecutive 24-hour periods, capturing different conditions (standing, eating, lying, handling). A Convolutional Neural Network (CNN) based tool was applied to develop a behavior classifier model using TensorFlow from the Keras 2.0.6 library. The labeled images were partitioned for training, validation and testing, and subjected to a cross-validation methodology.

Results: The final CNN-based model observed is composed of 3 convolution layers, each containing 64 kernels of size 5x5 and polling layer of size 2x2 (activation by Rectified Linear Unit). The convergence threshold was obtained in 10 epochs and the computational model tested with 569 frames (10%) obtained a general accuracy of 97.5%.

Discussion/Conclusion: Although the ‘standing’ class showed the highest number of incorrectly classified frames, in percentage terms in relation to the total frames of each class, the CNN-based model shows better performance for this class (precision of 98.4%). The precision (mean = 94.2%) and recall (mean = 94.7%) exhibited a more distinct outcome for the ‘eating’ class (83.3% and 87.7%, respectively) in comparison to the other classes. This indicates the smaller number of frames examples in the mentioned class influenced the model’s learning. The lower F1-Score value for the ‘eating’ class (0.85) compared to the other classes further supports this assertion. Less learning ability was identified to distinguish a horse standing close to the feeder region and a horse feeding, indicating that in the next stage of the work it is essential to expand the image database. In general, the results show the potential of the approach to compose a computer vision system for the studied application.

References

1. Greening, L., Downing, J., Amiouny, D., Lekang, L., McBride, S., 2021. The effect of altering routine husbandry factors on sleep duration and memory consolidation in the horse. Applied Animal Behaviour Science. https://doi.org/10.1016/j.applanim.2021.105229
2. Oliveira, T., Santos A., Silva J., Trindade P., Yamada A., Jaramillo F., Silva L., and Baccarin R.. . 2022. Hospitalisation and disease severity alter the resting pattern of horses. J. Equine Vet. Sci. 110:103832. doi: 10.1016/j.jevs.2021.103832.
3. Valletta, J.J., Torney, C., Kings, M., Thornton, A., Madden, J., 2017. Applications of machine learning in animal behaviour studies. Animal Behaviour 124, 203–220. https://doi.org/10.1016/j.anbehav.2016.12.005