Audio information retrieval for describing gait patterns in Brazilian horses

• Published in: Journal of animal science Vol. 98; p. 27
• Main Authors: Ventura, Ricardo V, Lopes, Rafael Z, Andrietta, Lucas T, Bussiman, Fernando, Balieiro, Julio, Carvalheiro, Roberto, Silva, Fabyano, Brito, Luiz, Alves, Anderson
• Format: Journal Article
• Language: English
• Published: Champaign Oxford University Press 01.11.2020
• Subjects: Animal breeding; Animal husbandry; Animals; Audio data; Audio signals; Data analysis; Economic sectors; Evaluation; Feature extraction; Gait; Horses; Information retrieval; Phenotypes; Principal components analysis; Quality control; Signal processing; Sound; Video data
• ISSN: 0021-8812; 1525-3163

The Brazilian gaited horse industry is growing steadily, even after a recession period that affected different economic sectors in the whole country. Recent numbers suggested an increase on the exports, which reveals the relevance of this horse market segment. Horses are classified according to the gait criteria, which divide the horses in two groups associated with the animal movements: lateral (Marcha Picada) or diagonal (Marcha_Batida). These two gait groups usually show remarkable differences related to speed and number of steps per fixed unit of time, among other factors. Audio retrieval refers to the process of information extraction obtained from audio signals. This new data analysis area, in comparison to traditional methods to evaluate and classify gait types (as, for example, human subjective evaluation and video monitoring), provides a potential method to collect phenotypes in a reduced cost manner. Audio files (n = 80) were obtained after extracting audio features from freely available YouTube videos. Videos were manually labeled according to the two gait groups (Marcha Picada or Marcha Batida) and thirty animals were used after a quality control filter step. This study aimed to investigate different metrics associated with audio signal processing, in order to first cluster animals according to the gait type and subsequently include additional traits that could be useful to improve accuracy during the identification of genetically superior animals. Twenty-eight metrics, based on frequency or physical audio aspects, were carried out individually or in groups of relative importance to perform Principal Component Analysis (PCA), as well as to describe the two gait types. The PCA results indicated that over 87% of the animals were correctly clustered. Challenges regarding environmental interferences and noises must be further investigated. These first findings suggest that audio information retrieval could potentially be implemented in animal breeding programs, aiming to improve horse gait.