A computer vision pipeline for bodyweight estimation in turkey hens

• Automated, non-contact body weight estimation using computer vision and machine learning system using a depth camera, offering a more efficient alternative to traditional, labor-intensive methods. • High predictive accuracy with up to 7% of absolute error across the animals’ lifetime, with input variables explaining 96% of the bodyweight variation. • Robust approach for commercial farm settings, as shown by the use of a "leave-one-animalout" data splitting approach, which emphasizes the model's ability to generalize to new, unseen birds. • Detailed feature engineering based on extracting unique three-dimensional features from the animals’ backs, showing a correlation as high as manually collected measurements. Efficient poultry growth is a critical factor in the industry's productivity and economic viability. Commercial farms routinely perform bodyweights on a portion of the flock to gather key insights to bird development, performance, and management. However, current approaches to collect bodyweights (BWs) are labor-intensive and time-consuming. Additionally, they fail to capture individual animal variation, since BWs are estimated by averaging the weights of birds measured together. This work proposes the use of a computer vision system based on depth cameras and machine learning (ML) algorithms for frequent, non-contact, and real-time monitoring of individual BWs. An observational study was conducted on a flock of 29 Nicholas Select turkey hens, monitored from 39 d to 105 d of age and manually weighed three times a week. Videos were captured at two frames per second using an Intel RealSense D435 camera, capturing color images overlapped with depth images of equal resolution, effectively providing visual and three-dimensional information from a top-down view over the animals’ back. A computational pipeline was developed to detect individual animals and extract three-dimensional high-level features from the captured images. An extensive experimental evaluation was conducted. Firstly, results focus on statistical characteristics of the dataset and the associations between BW and the extracted visual features. Then, a thorough evaluation of the computational pipeline is centered on the performance of several ML models, using grid search to find the optimal parameters, and using several data splitting approaches to explore various realistic commercial scenarios.