Audio-Based Characterization of Gait Parameters in Mangalarga Marchador, Campolina, and Piquira Horses Using Deep Learning.

Overview

• This study developed a deep learning method to analyze horse gait patterns using only audio recordings of hoof sounds.
• The method accurately estimates biomechanical gait parameters such as dissociation percentage in three Brazilian horse breeds, providing a practical, non-invasive tool for gait characterization.

Background and Motivation

• Gaited horses, especially those with a four-beat gait, have complex biomechanical movement patterns that are important for breed standards and performance assessments.
• Traditional gait evaluation methods often rely on subjective visual assessments or complex biomechanical analyses requiring specialized equipment, limiting accessibility and consistency.
• Objective, accessible, and non-invasive methods are needed to quantify gait parameters reliably.

Research Objective

• The primary goal was to use deep learning to estimate the dissociation percentage—a measure of the timing differences between hoof contacts in the four-beat gait—from audio signals alone.
• The study focuses on three Brazilian horse breeds known for their marcha gaits: Mangalarga Marchador, Campolina, and Piquira.

Data Collection and Preparation

• 268 audio samples were extracted from publicly available videos featuring the three horse breeds performing two specific gait types: marcha batida and marcha picada.
• Key acoustic features were extracted for each audio clip, including:
  • Root Mean Square Energy (RMS) – measures the intensity of the sound signal.
  • Zero-Crossing Rate (ZCR) – counts how often the audio signal crosses the zero amplitude line, related to the frequency content of the sound.
  • Mel-Frequency Cepstral Coefficients (MFCCs) – 13 coefficients representing the short-term power spectrum of the audio, widely used in audio and speech recognition.

Model and Methodology

• A Long Short-Term Memory (LSTM) neural network was chosen for its ability to model sequential data and temporal dependencies in audio signals.
• The model was trained with the extracted acoustic features to predict the time intervals between successive hoof-ground contacts.
• These predicted time intervals were then used to compute the dissociation percentage, a critical biomechanical parameter indicating the relative timing between hoof contacts.

Results

• The LSTM model showed very high predictive accuracy:
  • Correlation coefficient (R) between predicted and actual intervals was 0.98, indicating near-perfect linear agreement.
  • Mean Absolute Error (MAE) of 0.0071 seconds reflected very low prediction error.
• No significant differences in dissociation percentage were found between the two gait types (marcha batida and marcha picada), suggesting similar temporal hoof contact patterns.
• Significant breed-related differences were observed in:
  • Mean hoof-ground contact intervals.
  • Dissociation percentage values, reflecting distinct gait characteristics across breeds.
• Eight acoustic features were identified as particularly powerful in discriminating among breeds based on their gait audio signatures.

Conclusions and Implications

• The study demonstrates that biomechanical gait parameters in gaited horses can be accurately quantified from audio data alone without the need for visual or force-based sensors.
• This approach offers an objective, practical, and non-invasive tool for breeders, trainers, and researchers to analyze, standardize, and select for desirable gait characteristics.
• Beyond breed characterization, this method has potential applications in digital locomotion phenotyping, enabling the study of horse populations and genetic gait traits efficiently.
• The use of widely available video and audio data could facilitate large-scale, real-world gait analysis with minimal equipment and cost.

Future Directions

• Further validation on larger datasets and diverse environmental conditions would strengthen robustness and generalizability.
• Extension to other horse breeds and gait types could broaden applicability.
• Integration with mobile applications or wearable audio sensors might allow real-time gait monitoring in the field.
• Combining audio-based analysis with other sensor modalities could enhance multi-dimensional gait assessment.