To develop a robust non-contact image-based body weight prediction system, this study follows the experimental workflow shown below as its core framework, systematically conducting experimental design and data analysis. Depth cameras were used to collect image data of the cows' side, dorsal, and hip regions, and depth information from each region was processed accordingly. The data processing pipeline includes cow identification, missing value imputation, distance filtering, morphological operations, and connected component analysis to obtain more complete depth information of the cow's body surface and effectively remove extraneous depth noise. Subsequently, an image classifier was used to select valid, high-quality images from each region, and a calibration board area was used to establish an area fitting model for each region. Additional features include cow height, pixel counts per region, and the distance between each body region and the camera (mean and median values). Finally, a Feedforward Neural Network (FNN) and a Gaussian Process Regression (GPR) model were employed for weight prediction, and the prediction results were evaluated.
To ensure effective capture of cow images, this study adjusted the ground distance (D) and angle (θ) for each camera position. The ground distances for the side, dorsal, and hip cameras were D₁ = 212 cm, D₂ = 195.7 cm, and D₃ = 276.8 cm, respectively; the shooting angles were θ₁ = 50°, θ₂ = 90°, and θ₃ = 40°. In addition, the ground distances to the cross-beam beneath the barn eaves and to the eaves themselves were D₄ = 327 cm and D₅ = 382 cm. Using this hardware configuration, the study divided the electronic scale platform dimensions into coordinate points as a quantitative basis for removing the background from depth images and estimating the area features of each cow body region.
The figure below compares, for cow cid603 across three body regions (dorsal, hip, and side), the weekly actual body weight (green curve, from electronic scale measurements), FNN model predictions (blue curve), GPR model predictions (cyan curve), and farm-estimated weight (magenta curve, derived from body length and heart girth measurements). During the training phase (weeks 1–9), both FNN and GPR predictions closely followed the actual weight curve, demonstrating good training fit. The test phase (weeks 10–16, highlighted in lavender) was used to evaluate model generalisation performance.
Results show that during the test phase, both FNN and GPR models exhibited the smallest prediction errors in the dorsal region, with better prediction stability than in the hip and side regions. Notably, the GPR model's predictions in the dorsal region most closely tracked the actual weight trend. Although the GPR model performed slightly below the FNN on the overall training data, it demonstrated superior generalisation on the test dataset for non-pregnant cows. In contrast, the farm-estimated weights derived from body length and heart girth deviated markedly from measured values across all three regions, particularly in the test phase, where the curves showed large fluctuations and high errors, highlighting the limitations of traditional body measurement-based estimation methods.
