Automatic weight estimation of individual pigs using image analysis

作者：Kashiha Mohammadamin^*; Bahr Claudia; Ott Sanne; Moons Christel P H; Niewold Theo A; Oedberg Frank O; Berckmans Daniel

来源：Computers and Electronics in Agriculture, 2014, 107: 38-44.

DOI：10.1016/j.compag.2014.06.003

• Top-view body area
• Pig weight estimation
• Automated image processing
• Transfer function modelling
• Ellipse fitting

Health is a key element in pig welfare and steady weight gain is considered an indicator of good health and productivity. However, many diseases such as diarrhoea cause a substantial reduction in food intake and weight gain in pigs. Therefore, continuous weight monitoring is an essential method to ensure pigs are in good health. The purpose of this work was to investigate the feasibility of an automated method to estimate weight of individual pigs by using image processing.
This study comprised measurements on four pens of grower pigs, each consisting of 10 pigs. At the start of the experiments, pigs weighed on average 23 +/- 4.4 kg (mean +/- SD) while at the end their average weight was 45 +/- 6.5 kg. Each pen was monitored by a top-view camera. For validation purposes, the experiment was repeated once.
Individual pigs were automatically identified by their unique painting patterns using shape recognition techniques. The weight estimation process developed as follows: First, to localized pigs in the image, an ellipse fitting algorithm was employed. Second, the area the pig occupying in the ellipse was calculated. Finally, the weight of pigs was estimated using dynamic modelling. The developed model was then validated by comparing the estimated weight against manual twice weekly actual weight measurements of each individual pig. In addition, to monitor the weight of pigs individually, the pigs were marked on their back with basic unique paint patterns and were identified automatically using shape recognition techniques. In this way, the weight of each individual pig could be estimated. This method can replace the regular weight measurements on farms that require repeated handling and thereby causing stress to the pigs.
Overall, video imaging of fattening pigs appeared promising for real-time weight and growth monitoring. In this study pig weight could be estimated with an accuracy of 97.5% at group level (error of 0.82 kg) and 96.2% individually (error of 1.23 kg). This result is significant since the existing automated tools currently have a maximum accuracy of 95% (error of 2 kg) in practical setups and 97% (error of 1 kg) in walk-through systems (when pigs are forced to pass a corridor one by one) on average.
Future work should focus on developing specific algorithms to account for the effect of gender and genotype on body surface area and body weight since these factors affect the model parameters for weight estimation.

• 出版日期2014-9