Detecting Equine Gaits Through Rider-Worn Accelerometers.

Area of Science:

• Equine Science
• Biomechanical Engineering
• Machine Learning in Sports Science

Background:

• Direct sensor attachment for horse gait analysis poses risks like discomfort and entanglement.
• Rider-centric accelerometers offer a non-invasive alternative for horse movement classification.

Purpose of the Study:

• To evaluate the simultaneous impact of sensor location, sampling frequency, and window size on horse gait classification accuracy using rider-worn accelerometers.
• To identify the optimal parameters and model for accurate, non-invasive horse gait analysis.

Main Methods:

• Collected accelerometer data from four rider locations (knee, backbone, chest, arm) across five riders and seven horses performing three gaits (walk, trot, canter).
• Compared eight machine learning models, including a ConvLSTM2D network, analyzing data with varying sampling frequencies and four-second window sizes.
• Validated model performance using Leave One Subject Out Cross-Validation (LOSOCV).

Main Results:

• The ConvLSTM2D model achieved the highest average accuracy of 89.72% for classifying four horse gaits (halt, walk, trot, canter).
• Optimal performance was observed with a four-second interval width and a 25 Hz sampling frequency.
• An F1-score of 86.18% was achieved, demonstrating robust classification capabilities.

Conclusions:

• Rider-worn accelerometers provide an effective and non-invasive method for automatic horse gait classification.
• The ConvLSTM2D model, with optimized parameters, offers high accuracy for real-time equine movement analysis.
• This approach mitigates risks associated with direct sensor attachment, enhancing horse welfare and training insights.