Automatic lameness detection in sows

• Ano de defesa: 2024
• Autor(a) principal: Paula, Tauana Maria Carlos Guimarães de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizagem profunda; Artificial intelligence; Computer vision; Deep learning; Inteligência artificial; Locomoção; Locomotion; Pig; Suínos; Visão computacional
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/10/10134/tde-25092024-163531/

Lameness affects the animal′s locomotion, causing pain and discomfort and is a problem faced in pig farming with a prevalence in sows of up to 65% of animals. Observation of the animals goes unnoticed and identification of the sow\'s locomotion score is subjective as it depends on the experience of the evaluator. Computer vision systems detect sow lameness automatically, objectively, accurately and non-invasively. Locomotion has been studied using kinematics to extract complete movement data and the Fast Fourier Transform (FFT) to transform the time domain into the frequency domain. The objectives of this work were: to create a repository of images and videos of sows with different locomotion rates. Develop lateral and dorsal computer vision models to automatically identify and track key points on the sow\'s body, as well as detect different locomotion scores in sows. The video database was acquired from a commercial pig farm with the construction of a scenario for filming sows in locomotion with different lameness scores. Two cameras were used to record 2D videos. Thirteen locomotion experts evaluated the videos using the Locomotion Score System developed by Zinpro Corporation. From this annotated repository, computer models were trained and tested using the deep learning-based animal pose tracking framework SLEAP (Social LEAP Estimates Animal Poses). The best performing models were built using the LEAP architecture with 6 (lateral) and 10 (dorsal) skeletal keypoints. The architecture achieved average accuracy values of 0.90 and 0.72, average pixel distances of 6.83 and 11.37 and similarities of 0.94 and 0.86 for the lateral and dorsal views, respectively. Based on this, the kinematic study of the key points was used to develop the computer models and apply the FFT. The Naive Bayes, KNN, Random Forest and Multilayer Perceptron algorithms were tested with different methodologies and locomotion score conditions. The models that performed best in identifying sows with and without lameness were those tested with the Multilayer Perceptron algorithm. In the lateral view, the model with FFT achieved 91.5% accuracy; in the dorsal view, the model with kinematic data achieved 88.63% accuracy. Therefore, the computer models developed can be used to identify and track key points, as well as detect lameness in sows in the lateral and dorsal views with 2D images. In this way, they can contribute to the objective, precise and automatic evaluation of locomotion scores with the aim of improving the well-being, health and productivity of sows.