Preliminary analyses on the identification of horse behaviors using deep learning techniques

The observation of horse behavior offers significant insights into their subjective state, making it a valuable indicator of their welfare. Monitoring the time horses allocate to various behaviors throughout the day can serve as an effective strategy for assessing their well-being. Among these behaviors, sleep is particularly important due to its critical biological role in recovery and its cog- nitive function in memory consolidation. As prey animals, horses typically sleep for a total of 3 to 5 hours per day. Sleep deprivation is a common issue that acts as a major stressor, leading to alterations in both behavior and emotional state. Lying down has been recognized as a reliable indicator of welfare in stabled horses. Additionally, monitoring behaviors such as access to drinking water and feeding time can provide further in- sights into their welfare. However, directly observing these behaviors, whether in person or via video recordings, can be time-intensive, especially since horses spend only a small portion of their day lying down or drinking. To streamline daily management, computer vision technology offers automated methods to interpret and analyze visual data in animal environments. By leveraging techniques from image processing and machine learning, computer vision can extract valuable information and enhance the understanding of animal behaviors. This study investigates the use of a deep learning-based computer vision system to identify the behaviors of individual stabled horses. The initial step involved fine-tuning a pre-trained YOLO architecture to recognize specific behaviors, such as lying, active standing, non-active standing, and drinking, for a single horse housed in an enclosed box. Object detection methods were em- ployed to identify lying and standing behaviors, while pose estimation techniques were used to detect drinking activity. To distinguish between active and non-active standing, a pixel-based threshold was applied. The system was then utilized for continuous monitoring over one month, generating a 24-hour time budget for the horse. The performance of the model was evaluated using precision-recall curves and by comparing its behavior classifications with manual annotations of the same video data. The system demonstrated an 86% accuracy in behavior identification relative to human labeling. The technology presented in this study enables real-time recognition and provides valuable information on the welfare of monitored animals. The results highlight the potential of this approach for enhancing the monitor- ing and understanding of horse behavior.