Non‐Invasive and Low‐Cost Approach for Estimating Sheep Weight Using Computer Vision

Traditional methods for sheep weighing are invasive, stressful, and often impractical for small and medium‐sized producers. While advances in computer vision and artificial intelligence enable non‐invasive weight estimation, implementing these techniques in real‐world agricultural scenarios still faces the challenge of requiring robust and scalable solutions. We developed a video‐based system for non‐invasive sheep weight estimation using a Microsoft Kinect V2 sensor. The methodology integrates automated frame selection with the EfficientNetB3 model, segmentation with a U‐Net architecture featuring a DenseNet169 backbone, and weight estimation via convolutional neural networks (CNNs) and an ensemble approach to enhance accuracy. The EfficientNetB3 model achieved the highest performance in frame selection, with an F1‐score of 95.43%. The segmentation model obtained a Dice coefficient of 96% and a Jaccard index of 92%, ensuring precise animal isolation. Weight estimation using EfficientNetB2 combined with linear regression at a resolution of 1024 × 1024 pixels resulted in a Root Mean Square Error (RMSE) of 3.44 kg and a coefficient of determination ( R 2 ) of 0.90. The ensemble approach further improved accuracy, reducing the RMSE to 3.44 kg and increasing the R 2 to 0.90. The proposed system provides a non‐invasive, automated, and cost‐effective solution for sheep weight monitoring that minimizes human intervention and reduces animal stress. Future research will focus on refining segmentation techniques and optimizing computational efficiency for broader applicability.