Recognition of horse behaviours using deep learning techniques

The observation of horse behaviour offers important understanding into their state, making it a key indicator of their welfare. Among these be-haviours, sleep is particularly important due to its critical biological role in recovery and its cogni-tive function in memory consolidation [1]. As-sessing lying behaviour is an essential aspect of equine welfare evaluation, as horses generally tend to sleep shortly after lying down [3]. As prey animals, horses typically spend between 4 to 15 hours per day to standing rest, while the time spent lying down can range from minutes to several hours [2]. On average, adult horses spent about 80% of their resting time standing [4]. Moreover, monitoring additional behaviours such as access to drinking water and feeding time can provide further information into their welfare. However, di-rectly observing these behaviours, whether in per-son or via video recordings, can be time-consuming, especially since horses spend only a small portion of their day lying down or drinking. To improve daily management, computer vision technology offers automated methods to interpret and analyze visual data in animal environments [5]. Utilizing methods from image processing and machine learning, computer vision can extract meaningful data and improve the understanding of animal behaviours. This study investigates the use of a deep learning-based computer vision system to identify the behaviours of individual stabled horses. The initial step involved fine-tuning a pre-trained YOLO architecture to recognize specific behaviours, such as lying, active standing, non-active standing, and drinking, for a single horse housed in a closed box. Object detection methods were used to identify lying and standing behaviours, while pose estimation techniques were utilized to detect drinking activity. To differentiate 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 behaviour classifications with manual annotations of the same video data. The system demonstrated an 86% accuracy in behaviour identification relative to human labelling. In conclusion, the technology presented in this study allows real-time recognition and pro-vides valuable information on the welfare of monitored animals. The results highlight the potential of this approach for improving the monitoring and understanding of horse behaviour.