Module 101 / 03
Lameness monitoring
- 1Two cameras in the alley
- 238 keypoints
- 3Joint-angle dynamics
- 4Daily lameness index
A project of the Artificial Intelligence Lab, Timiryazev University
AgroBioSense
Multimodal AI platform for cattle health monitoring and digital phenotyping
Depth cameras and computer vision turn video of an animal into a measurable digital profile — gait, joint angles and conformation measurements — without touching the animal or stopping production.
Gold medal · International Exhibition of Inventions, Guangzhou 2026
- anatomical landmarks in the conformation model
- 43
- animals per day vs. 24 with manual scoring
- up to 480
- agreement with expert assessment on 150 animals
- >90%
- to capture one animal
- 3–5 s
Demo
The system at work
Dashboard recordings from real farms: lameness monitoring, linear conformation assessment and detection at the feed bunk.
Module 202 / 03
Linear conformation assessment
- 1RealSense D435 RGB-D camera
- 243 anatomical landmarks
- 37 measurements in centimetres
- 4Overall score on a 1–9 scale
Roadmap · behaviour03 / 03
Feed bunk behaviour
- 1Feed bunk zone
- 2Head detection
- 3Feeding behaviour
The challenge
From occasional inspection to continuous monitoring
Lameness and conformation are still assessed visually and occasionally: specialists react once signs are obvious, and results depend on their experience and viewing conditions.
Conventional approach
- 1Clinical sign
- 2Inspection
- 3Diagnosis
- 4Treatment
Response after symptoms become pronounced
AgroBioSense approach
- 1Continuous RGB-D data stream
- 2Pattern change
- 3Risk signal
- 4Targeted veterinary check
Early detection of deviations before clinical signs
- to score one animal manually
- up to 30 min
- agreement of visual scoring on complex traits
- up to 70%
- occasional assessments per production cycle
- 1–2
How it works
From video stream to a digital animal profile
A complete data cycle — from image capture to a signal for the veterinarian.
- 01
Data capture
RGB-D cameras record the animal from several angles; RFID links every frame to a specific animal.
- 02
Detection and tracking
A neural network finds and follows each animal in the video stream.
- 03
Pose estimation
The model locates anatomical landmarks: 38 keypoints for locomotion and 43 for conformation.
- 04
3D reconstruction
The depth map converts keypoint coordinates into metric space.
- 05
Trait calculation
Body measurements, joint angles, symmetry and indices in centimetres and degrees.
- 06
Time series
Each animal's individual baseline and deviations from it.
- 07
Risk signal
A real-time dashboard and a priority list of animals to inspect.
Outcome
A digital animal profile as the basis for proactive veterinary care
Platform modules
Two modules, one technology core
Both modules share the same RGB-D capture and pose-estimation pipeline but solve different tasks: veterinary care and breeding.
Module 1
Limb Health and Biomechanics Monitoring
The network does not just see “lame / not lame” — it sees changes in movement mechanics. A veterinarian may not yet notice obvious lameness while the structure of the animal's movement has already changed.
- YOLOv8l-pose model with 38 anatomical keypoints
- Coefficient of variation of joint angles: 2–3.4% in healthy animals, 19.2% in early lameness, 21.8–26.3% in clinical lameness
- Lameness index and joint-angle dynamics for every animal in the dashboard
- recall
- 91.74%
- keypoint mAP@0.5
- 83.11%
- accuracy in detecting locomotion disorders
- up to 96.2%
- annotated images of 449 cows
- 12,000
Module 2
Contactless Linear Conformation Assessment System
The system reports not only a final score but the measured trait itself in centimetres and degrees, so results can be re-checked, compared across animals and tracked over time.
- Stature, body depth, chest width, rump angle and width, rear legs side view, udder depth
- Dedicated udder morphometry: 16 reference points and 8 traits (in extended validation)
- Measurements converted to the 1–9 linear scoring scale and a digital record for breeding work
- anatomical landmarks
- 43
- conformation traits in the current version
- 7
- agreement with expert instrumental assessment
- >90%
- higher throughput
- ×20
Results
Pilot deployment and impact
In 2026 the system passed pilot production testing at livestock farms.
- RGB-D recording passes, 3,534 synchronised streams
- 1,178
- project archive of production footage
- >100 TB
Herd of 1,000 animals: manual scoring vs. digital pipeline
| Metric | Manual scoring | AgroBioSense |
|---|---|---|
| Throughput | up to 24 animals a day | up to 480 animals a day |
| Full assessment cycle | ≈ 41.7 working days | ≈ 2.1 working days |
| Time per animal | up to 30 minutes | 3–5 seconds of capture plus automatic processing |
| Output | specialist's score | measurements in cm and degrees, evidence frames, 1–9 score |
| Scalability | limited by the number of classifiers and site visits | mass phenotyping with digital records |
Ekopole LLC
Verkhny Mamon, Voronezh region
RGB-D data collection under normal production conditions, camera calibration, landmark localisation testing and dashboard output.
Belgorod State Agrarian University
Belgorod
Joint research on detecting animals' eyes in farm imagery.
Awards
Recognition
Gold medal
12th International Exhibition of Inventions
China Association of Inventions
Guangzhou, China · 21–23 August 2026
Invention “Artificial intelligence for early diagnosis of lameness and linear assessment of cattle's exterior”. Certificate No. 202646RU13603.
Certificate, PDFWinner
Gravitatsiya-2026
International University Award in Artificial Intelligence and Big Data
2026
Category “Algorithms and Software Solutions in AI and Big Data”. Project “AI platform for digital phenotyping of cattle to analyse biomechanical health patterns and linear conformation traits”.
Certificate, PDFCertificates
State registration certificates
Key results are protected by Rospatent certificates of state registration for computer programs and databases. Right holder: Russian State Agrarian University — Moscow Timiryazev Agricultural Academy.
- DatabaseNo. 2026622755, 16 June 2026
Multimodal RGB-D imaging database of Holstein cattle for contactless live-weight estimation
Open PDF· in Russian - Computer programNo. 2026666742, 3 June 2026
Program for adaptive selection of informative video frames to localise landmarks in cattle conformation assessment
Open PDF· in Russian - DatabaseNo. 2026622314, 22 May 2026
Cattle image dataset for locomotor function analysis and morphological assessment based on 38 skeletal keypoints
Open PDF· in Russian - DatabaseNo. 2025626469, 24 December 2025
Cow skeletal keypoints for keypoint detection
Open PDF· in Russian - Computer programNo. 2025695196, 10 December 2025
Neural network for early-stage lameness classification
Open PDF· in Russian - Computer programNo. 2025695364, 10 December 2025
Program for statistical analysis and visualisation of cattle joint biomechanics for lameness diagnosis
Open PDF· in Russian
Publications
Publications and talks
- 2026Article
Automated Dairy Cattle Lameness Assessment Using a 38-Keypoint Whole-Body YOLOv8l-Pose Model
Veterinary World (Q1)
Accepted for publication
- 2026Talk
Livestock: automating health-pattern monitoring with computer vision
2nd International Forum “Veterinary Safety”, Ministry of Agriculture of the Russian Federation
23–25 September 2026
Team
Project team
Anastasia Grecheneva
Project lead, PhD (Engineering), Associate Professor, Director of the Project Institute for Digital Transformation of Agriculture
Dmitry Proshin
Maxim Baknin
Sergey Lapshin
Sergey Akchurin
Alexander Braginets
Nikita Farafonov
Nikita Vlasyuk
Partners
Timiryazev University (RSAU–MTAA)
Developer and right holder
Ekopole LLC
Pilot production site
Belgorod State Agrarian University
Joint research
What's next
Platform roadmap
The same technology core is being extended to new traits and species. These directions are under research and are not part of the current modules.
Behaviour monitoring
Activity, feed bunk visits and lying time as digital biomarkers compared with each animal's own baseline.
Integrated risk pattern
A combined score from behaviour, morphometry and locomotion, including indirect signs of metabolic disorders. The system flags animals for inspection; it does not diagnose.
Other species
Pigs, poultry and companion animals, with species-specific keypoints and separate clinical validation.
Herd management integration
Exchanging structured animal records with herd management systems; the data format is designed for Russia's federal breeding resources information system.
Contact
Discuss a pilot or partnership
Tell us about your farm or research task — the project team will reply by email.
Or email us
info@timailab.ru