Monitoring Horses in Stalls: From Object to Event Detection
Authors: Dmitrii Galimzianov, Viacheslav Vyshegorodtsev, Ivan Nezhivykh
Journal: ArXiv
Summary
# Editorial Summary: Automated Behavioural Monitoring in Stabled Horses Real-time monitoring of stalled horses remains practically challenging despite its critical importance for early identification of lameness, colic, stereotypies, and other welfare concerns; manual observation is labour-intensive and inconsistent across facilities. Researchers developed a vision-based system combining YOLOv11 object detection with BoT-SORT multi-object tracking to automatically identify horses, personnel, and behavioural events within stable environments, with event classification derived from trajectory analysis and spatial relationships rather than raw video frames alone. The prototype successfully distinguished five event types and demonstrated reliable detection of horse-related behaviours whilst acknowledging that people detection remained limited by insufficient training data—a recognised constraint in agricultural computer vision applications. Blind spots created by camera positioning were systematically mapped and accounted for in the system architecture, addressing a practical concern often overlooked in surveillance studies. For practitioners seeking to implement automated monitoring systems, this foundation suggests considerable potential for objective welfare assessment and management optimisation, though further refinement of multi-class detection accuracy and validation across diverse stable designs will be necessary before widespread clinical or commercial deployment.
Read the full abstract on the publisher's site
Practical Takeaways
- •Automated stall monitoring systems can reduce labor-intensive manual observation and enable early detection of behavioral changes indicative of illness or welfare concerns
- •Camera placement and blind spot compensation are critical considerations for practical implementation in real stable environments
- •Current technology reliably tracks horse behavior but may require additional training data for robust personnel detection in busy facilities
Key Findings
- •A vision-based monitoring system using YOLOv11 and BoT-SORT successfully detected and tracked horses and personnel in stalls with reliable performance for horse-related events
- •Custom dataset construction was supported by foundation models CLIP and GroundingDINO to annotate stall behavior
- •System distinguished five event types and accounted for camera blind spots, though people detection was limited by data scarcity
- •Automated behavioral monitoring shows potential for early detection of health and welfare issues without continuous human observation