Automatic assess of growing-finishing pigs\' weight through depth image analysis

• Ano de defesa: 2017
• Autor(a) principal: Condotta, Isabella Cardoso Ferreira da Silva
• 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: Kinect® sensor; Pesagem; Precision livestock farming; Sensor Kinect ® Weighing; Zootecnia de precisão
• Link de acesso: http://www.teses.usp.br/teses/disponiveis/11/11152/tde-03082017-093143/

A method of continuously monitoring weight would aid producers by ensuring all pigs are gaining weight and increasing the precision of marketing pigs thus saving money. Electronically monitoring weight without moving the pigs to the scale would eliminate a stress-generating source. Therefore, the development of methods for monitoring the physical conditions of animals from a distance appears as a necessity for obtaining data with higher quality. In pigs\' production, animals\' weighing is a practice that represents an important role in the control of the factors that affect the performance of the herd and it is an important factor on the production\'s monitoring. Therefore, this research aimed to extract weight data of pigs through depth images. First, a validation of 5 Kinect &reg; depth sensors was completed to understand the accuracy of the depth sensors. In addition, equations were generated to correct the dimensions\' data (length, area and volume) provided by these sensors for any distance between the sensor and the animals. Depth images and weights of finishing pigs (gilts and barrows) of three commercial lines (Landrace, Duroc and Yorkshire based) were acquired. Then, the images were analyzed with the MATLAB software (2016a). The pigs on the images were selected by depth differences and their volumes were calculated and then adjusted using the correction equation developed. Also, pigs\' dimensions were acquired for updating existing data. Curves of weight versus corrected volumes and corrected dimensions versus weight were adjusted. Equations for weight predictions through volume were adjusted for gilts and barrows and for each of the three commercial lines used. A reduced equation for all the data, without considering differences between sexes and genetic lines was also adjusted and compared with the individual equations using the Efroymson\'s algorithm. The result showed that there was no significant difference between the reduced equation and the individual equations for barrows and gilts (p<0.05), and the global equation was also no different than individual equations for each of the three sire lines (p<0.05). The global equation can predict weights from a depth sensor with an R2 of 0,9905. Therefore, the results of this study show that the depth sensor would be a reasonable approach to continuously monitor weights.