2026-09-10 ペンシルベニア州立大学（Penn State）
＜関連情報＞
- https://www.psu.edu/news/research/story/camera-ai-assess-turkeys-predict-future-body-weight-study
- https://www.frontiersin.org/journals/animal-science/articles/10.3389/fanim.2026.1868845/full
コンピュータービジョンを用いた雄七面鳥の体重および体長の予測 Computer vision-based forecasting of body weight and carcass in tom turkeys
Mireia Molins,Enrico Casella
Frontiers in Animal Science Published:19 June 2026
DOI:https://doi.org/10.3389/fanim.2026.1868845
Abstract
Predicting future body weight (BW) trajectories in poultry production offers substantial advantages for operational efficiency, nutritional management optimization, and environmental sustainability. Traditional approaches to individual BW monitoring require extensive manual labor and frequent animal handling, creating both economic and welfare challenges. Although recent advances in computer vision (CV) and artificial intelligence have demonstrated capability in real-time BW estimation, the application of these technologies for temporal forecasting, i.e., predicting weight at future time points, remains unexplored. This investigation implemented a longitudinal monitoring system to assess whether CV-based methods could generate reliable forward-looking BW predictions in tom turkeys. An Intel RealSense D435 cameras was used to capture overhead footage of overlapped color and depth images of 30 tom turkeys housed collectively and observed from day 37 of age through day 133. Reference weights were obtained through manual weighing five times weekly. After isolating individual birds using instance segmentation, ResNet architectures were trained to perform same-day BW estimation, BW forecast up to three weeks in the future, and chilled carcass weight predictions. Models were fine-tuned using a leave-one-animal-out cross validation, despite the significant computational overhead associated with deep learning iterations. Findings indicate strong predictive performance: same-day BW estimation achieved an R² of 0.96 ± 0.04 with 7.39 ± 4.06% mean absolute percentage error (MAPE). The temporal forecasting models maintained robust accuracy, yielding MAPE 7.15 ± 3.17%, while chilled carcass yield prediction the model achieved MAPE of 7.53 ± 4.30%, showing the technical feasibility of AI-driven BW forecasting in poultry production systems.
