End-to-End Horse Gait Classification in Uncontrolled Environments Using Inertial Sensors
Authors: Mahaut Gérard, S. Hanne-Poujade, Guillaume Dubois, Henry Chateau, N. Mezghani
Journal: IEEE Access
Summary
# Editorial Summary: Automated Gait Classification Using Inertial Sensors Objective gait identification is fundamental to lameness assessment, yet current clinical practice relies on subjective visual evaluation at predetermined gaits. Researchers from the IEEE Access journal developed an automated classification system using seven inertial measurement units (IMUs) positioned across the horse's limbs, head, withers, and pelvis, processing raw sensor data without manual signal pre-selection—an approach that more closely mimics real-world diagnostic scenarios than previous methods. Testing three algorithmic approaches (XGBoost machine learning, LSTM deep learning, and transfer learning via ENCOD-CNN) across 110 labelled horses, the transfer learning model achieved 91.9% accuracy for four-gait classification and 97.1% accuracy specifically for gallop identification, with the complete pipeline delivering 91.2% overall accuracy. Critically, the study quantified the training dataset size required for robust performance, using 1440 horses in an unsupervised model to establish baseline parameters. For practitioners, this work demonstrates that IMU-based automated gait classification is clinically viable and could standardise lameness detection across different operators and environments, particularly valuable for identifying asymmetries during trot—the diagnostic gait of choice—whilst reducing the subjective variability that currently limits precision in equine orthopaedic assessment.
Read the full abstract on the publisher's site
Practical Takeaways
- •IMU-based systems can objectively classify gaits with >91% accuracy in real clinical settings, potentially improving objective lameness assessment beyond visual evaluation
- •Veterinarians using these systems need only ~110 labeled reference cases to train effective models, making implementation practical for clinical adoption
- •Automated gait identification during examination allows consistent comparison of vertical displacement symmetry across different horses and populations
Key Findings
- •Transfer learning (ENCOD-CNN) achieved 91.9% accuracy for four-gait classification using raw IMU data from seven strategically placed sensors
- •Gallop-specific classification achieved 97.1% accuracy with the integrated pipeline reaching 91.2% overall accuracy
- •End-to-end gait classification is feasible in uncontrolled environments without manual signal segmentation, using 110 labeled horses from a dataset of 1440