Horse gaits classification using the Infinite Gaussian Mixture Model and stability analysis

[Speaker] Ohkita, Midori:1
[Co-author] Nishiyama, Keita:2, Mano, Hiroshi:3, Murai, Chizuko:4, Takagi, Tokio:5, Kubo, Takatomi:6, Ikeda, Kazushi:6, Sawa, Kosuke:1, Samejima, Kazuyuki:4
1:Senshu University (Japan), 2:Teikyo University of Science (Japan), 3:EverySense,Inc (Japan), 4:Tamagawa University (Japan), 5:Tokyo University (Japan), 6:Nara Institute of Science and Technology (Japan)

Horse gaits (walk, trot, and canter) classification by human judge is subjective and difficult to precisely detect gait transition. Therefore, we used a machine learning method for gaits classification from acceleration data. The acceleration data was converted into a short time spectrum by Fourier transformation, and probabilistic classification by the Infinite Gaussian Mixture Model (IGMM) was applied to the compressed dimensions of the spectrum. The matching rate between the classes classified by IGMM and that by judge reached 94.98 ± 1.37% (mean ± SD). IGMM detected slight behavioral changes that could not be detected by judge. The time resolution of detecting unstable dynamics of acceleration during gait transition by the stability analysis using a Poincaré map was about 0.33s. This precise classification could be used to investigate how human signals are transmitted to horses at short intervals in horseback riding regarded as one of social interactions between humans and animals.
Advanced Search