Preferência de galinhas para diferentes comprimentos de onda de luz usando deep learning
Carregando...
Data
2022-01-28
Autores
Rodrigues Siriani, Allan Lincoln [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
O objetivo deste trabalho foi desenvolver um modelo de rede neural convolucional utilizando Deep Leaning e Framework YOLO para identificar e localizar galinhas em vídeos gravados em ambientes com baixa iluminação e pouco contraste e desenvolver um algoritmo para rastrear a movimentação das aves nessas condições. O modelo foi gerado e posteriormente aplicado em vídeos de um experimento de comportamento de preferência de galinhas da linhagem H&N Brown Nick por três ambientes de iluminação monocromática (branco, vermelho e verde). As aves tinham livre acesso e trânsito a esses compartimentos. Além dos tratamentos de iluminação, foram testados três tratamentos de temperatura: calor (35ºC), termoneutro (24ºC) e frio (17ºC). Os vídeos foram gravados em baixa iluminação (~5 lx) com auxílio de fonte de iluminação near infrared da própria câmera, resultando em imagens em escala de cinza de baixa qualidade e baixo contraste das aves com o restante do ambiente. O modelo gerado apresentou resultados de 99,9% de acurácia para a identificação e localização das aves nos compartimentos da câmara de preferência. A partir da identificação das aves, foi possível identificar o tempo de permanência das aves em 648 horas de vídeo. A análise dos resultados indicou que as galinhas têm preferência em permanecer no ambiente com luz branca, seguida da vermelha e verde. Na condição de calor foi observado uma distribuição mais igualitária das aves entre os compartimentos, de modo que essa dispersão das aves é indício de que a condição térmica sobrepõe a preferência das aves pelo ambiente luminoso. A partir da localização das aves, foi desenvolvido um algoritmo para análise da movimentação (rastreamento) das aves usando filtro de Kalman, onde foi possível seguir as aves individualmente e medir a distância percorrida de cada indivíduo. As técnicas de visão computacional aplicadas nesta pesquisa podem vir a ser aperfeiçoadas e aplicadas futuramente em condições comerciais, contribuindo para a melhoria da produção e bem-estar das aves.
The objective of this work was to develop a convolutional neural network model using Deep Leaning and YOLO framework to identify and locate chickens in videos recorded in low-light and low-contrast environments and to develop an algorithm to track the movement of birds in these conditions. The model was generated and later applied to videos of a preference behavior experiment of H&N Brown Nick hens by three monochromatic lighting environments (white, red and green). The birds had free access and transit to these compartments. In addition to the lighting treatments, three temperature treatments were tested: heat (35ºC), thermoneutral (24ºC) and cold (17ºC). The videos were recorded in low lighting (~5 lx) with the aid of the camera's own near infrared light source, resulting in low quality grayscale images and low contrast of the birds with the rest of the environment. The generated model presented results of 99.9% of accuracy for the identification and location of birds in the compartments of the chamber of preference. From the identification of the birds, it was possible to identify the length of stay of the birds in 648 hours of video. The analysis of the results indicated that the chickens prefer to remain in the environment with white light, followed by red and green. In the heat condition, a more egalitarian distribution of birds between the compartments was observed, so that this dispersion of birds is an indication that the thermal condition overrides the preference of birds for the bright environment. From the location of the birds, an algorithm was developed to analyze the movement (tracking) of the birds using a Kalman filter, where it was possible to follow the birds individually and measure the distance traveled by each individual. The computer vision techniques applied in this research can be improved and applied in the future under commercial conditions, contributing to birds' production improvement and welfare.
The objective of this work was to develop a convolutional neural network model using Deep Leaning and YOLO framework to identify and locate chickens in videos recorded in low-light and low-contrast environments and to develop an algorithm to track the movement of birds in these conditions. The model was generated and later applied to videos of a preference behavior experiment of H&N Brown Nick hens by three monochromatic lighting environments (white, red and green). The birds had free access and transit to these compartments. In addition to the lighting treatments, three temperature treatments were tested: heat (35ºC), thermoneutral (24ºC) and cold (17ºC). The videos were recorded in low lighting (~5 lx) with the aid of the camera's own near infrared light source, resulting in low quality grayscale images and low contrast of the birds with the rest of the environment. The generated model presented results of 99.9% of accuracy for the identification and location of birds in the compartments of the chamber of preference. From the identification of the birds, it was possible to identify the length of stay of the birds in 648 hours of video. The analysis of the results indicated that the chickens prefer to remain in the environment with white light, followed by red and green. In the heat condition, a more egalitarian distribution of birds between the compartments was observed, so that this dispersion of birds is an indication that the thermal condition overrides the preference of birds for the bright environment. From the location of the birds, an algorithm was developed to analyze the movement (tracking) of the birds using a Kalman filter, where it was possible to follow the birds individually and measure the distance traveled by each individual. The computer vision techniques applied in this research can be improved and applied in the future under commercial conditions, contributing to birds' production improvement and welfare.
Descrição
Palavras-chave
Análise de imagens, Estresse térmico, Comportamento animal, Filtro de Kalman, YOLO, Zootecnia de precisão, Image analysis, Machine learning, Animal behavior, Kalman filter, Precision Livestock Farming