Monitoring Horses in Stalls: From Object to Event Detection

summary

The gist

The gist: This study presents a prototype vision-based monitoring system that automates the detection and tracking of horses and people inside stables using object detection and multi-object tracking

In short

This study created a prototype vision system using YOLOv11 and BoT-SORT to automatically detect and track horses and people in stables. The system successfully detected horse-related events with high accuracy but struggled to reliably detect people due to limited training data. It establishes a foundation for real-time equine behavioral monitoring.

Key concepts

YOLOv11: This is a state-of-the-art object detection model used by the system to identify and locate horses and people within video frames. It scans images to draw boxes around objects, helping the computer understand what is present in the stable environment.
BoT-SORT: This algorithm is used for multi-object tracking. Once an object is detected, BoT-SORT keeps track of it across consecutive frames, assigning it a unique ID. This helps the system follow a horse or person as they move through the video.
Event Detection System: This component converts tracked object movements into meaningful events, such as a horse entering or leaving a stall. It analyzes trajectories and spatial relationships to determine when specific behaviors occur, allowing for automated logging of activities.
CLIP and GroundingDINO: These are foundation models used to help create the custom dataset by assisting with image annotation. They help speed up the process of labeling images, making it easier to train the detection models accurately.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Federated Learning — explained in 17 episodes
Retrieval-Augmented Generation (RAG) — explained in 16 episodes
Interpretability — explained in 15 episodes
Quantum Fisher Information (QFI) — explained in 15 episodes
Sample Complexity — explained in 15 episodes

This episode discusses

The paper

Monitoring Horses in Stalls: From Object to Event Detection · Read on arXiv

Dmitrii Galimzianov, Viacheslav Vyshegorodtsev, Ivan Nezhivykh

Monitoring the behavior of stalled horses is essential for early detection of health and welfare issues but remains labor-intensive and time-consuming. In this study, we present a prototype vision-based monitoring system that automates the detection and tracking of horses and people inside stables using object detection and multi-object tracking techniques. The system leverages YOLOv11 and BoT-SORT for detection and tracking, while event states are inferred based on object trajectories and spatial relations within the stall. To support development, we constructed a custom dataset annotated with assistance from foundation models CLIP and GroundingDINO. The system distinguishes between five event types and accounts for the camera's blind spots. Qualitative evaluation demonstrated reliable performance for horse-related events, while highlighting limitations in detecting people due to data scarcity. This work provides a foundation for real-time behavioral monitoring in equine facilities, with implications for animal welfare and stable management.

DOI: 10.1134/S1064562425700437

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Monitoring Horses in Stalls: From Object to Event Detection".

Jane: The gist: This study presents a prototype vision-based monitoring system that automates the detection and tracking of horses and people inside stables using object detection and multi-object tracking techniques.

Tom: First, who's behind it and why it matters.

Paper summary: Jane: So we’ve covered how they built this whole system for monitoring horses and people in stables, starting with object detection and ending with event tracking using YOLOv11 and BoT-SORT <ref:2510.17409#pg2>.

Lu: The paper itself is titled "Monitoring Horses in Stalls: From Object to Event Detection," which really tells you the journey they took from just seeing a horse to actually understanding what's happening in the stall <ref:2510.17409#pg1>.

Tom: The authors did do a qualitative analysis, which showed they’re quite confident about the horse events but also flagged that detecting people is tricky because there just aren't enough examples of them in their training data <ref:2510.17409#pg3>.

Meng: So for a startup or someone managing a farm, what does this actually mean when you look at the practical side? Does it mean we can put this on a camera system right away?

Jane: It means they laid the groundwork for real-time monitoring, which is huge for animal welfare and how stables get managed, even if they know they have to focus on getting more data for the people part <ref:2510.17409#pg3>.

Lu: The main point of this work is showing that you can use these computer vision techniques to estimate when something starts or stops happening based on where it moves around in space <ref:2510.17409#pg3>.

Lalam: It shows how improving the AI models, especially for those harder classes like people, can help build better systems for tracking and understanding complex behaviors within any environment <ref:2510.17409#pg3>.

Tom: That leads us right into the next part of this research—how they plan to actually make that person detection class much stronger with more data.

Jane: The authors developed a prototype system for monitoring horses and people in stables using computer vision techniques <ref:2510.17409#pg3>.

Lu: They successfully employed YOLOv11 and BoT-SORT for detection and tracking, while inferring event states based on object trajectories and spatial relations within the stall <ref:2510.17409#pg2>.

Meng: The paper lays out that this system distinguishes between five event types and accounts for the camera's blind spots <ref:2510.17409#pg1>.

Tom: So overall, it’s a solid piece of work that proves you can use these tools to automate monitoring, even if the current data set for people is lacking <ref:2510.17409#pg3>.

Jane: The main point of this work is showing that you can use these computer vision techniques to estimate when something starts or stops happening based on where it moves around in space <ref:2510.17409#pg3>.

Lu: And they acknowledge that the temporal Intersection over Union metric is a suitable measure for evaluating a system's ability to detect event start and end times <ref:2510.17409#pg3>.

Meng: It shows how improving the AI models, especially for those harder classes like people, can help build better systems for tracking and understanding complex behaviors within any environment <ref:2510.17409#pg3>.

Tom: That’s the path forward, building on this prototype to make a more robust system that can handle both horses and people effectively in real-time environments <ref:2510.17409#pg3>.

Conclusion: Tom: So, we've been talking about how they built this whole system for monitoring horses and people in stables, starting with object detection and ending with event tracking using YOLOv11 and BoT-SORT.

Jane: The paper itself is titled "Monitoring Horses in Stalls: From Object to Event Detection," which really tells you the journey they took from just seeing a horse to actually understanding what's happening in the stall.

Lu: The authors did do a qualitative analysis, which showed they’re quite confident about the horse events but also flagged that detecting people is tricky because there just aren't enough examples of them in their training data.

Meng: So for a startup or someone managing a farm, what does this actually mean when you look at the practical side? Does it mean we can put this on a camera system right away?

Tom: It means they laid the groundwork for real-time monitoring, which is huge for animal welfare and how stables get managed, even if they know they have to focus on getting more data for the people part.

Jane: The main point of this work is showing you can use these computer vision techniques to estimate when something starts or stops happening based on where it moves around in space.

Tom: That leads us right into the next part of this research—how they plan to actually make that person detection class much stronger.

More episodes

← Home