Body Condition Scoring of Dairy Cows Based on Feature Point Location

The use of computer vision to estimate the Body Condition Score for cow has demonstrated to be feasible. However, most research has been limited to fixed camera positions, which restricts the technique’s usefulness. This research acquired cow data at various distances and angles to investigate the impact of distance and different depth images restoration method on scoring accuracy. A U-Net neural network-based model was employed for background segmentation. The model proposed in this study performs best among SegNet, UNet, UNet++ and DeeplabV3, with significant performance improvement and superior overall segmentation accuracy, and has stronger foreground object identification capability. Additionally, we utilize the concept of human feature point localization to pinpoint the positions of cow feature points. The results show that compared to Hourglass, CPN, and Hrnet, the model in this study has significant advantages in three core indicators: accuracy, recall, and mAP. Moreover, we presented an unsupervised depth image reconstruction model based on the Denoising Diffusion Probabilistic Model and Unet++ Mode, facilitating the measurement and scoring of cow characteristics. Finally, the measurement results of cow body size using the Lerp model, autoregressive model, GAN model, and the depth image completion model proposed in this study were compared. The method proposed in this study was found to be effective and feasible for meeting actual production requirements at a camera distance of 1–2 m from the cow, achieving high accuracy with a coefficient of determination above 0.9. Additionally, the three models used were effective in handling small-scale depth deficits, with a high degree of agreement between machine and manual measurements. The DDPM-based depth image reconstruction model was found to produce the most accurate results when the camera distance from the cow was between 1-3m. Therefore, this research makes it possible to make flexible measurements in the near range using existing methods. Accurate measurements at longer ranges are influenced by depth image quality and feature point localization accuracy, which require more advanced techniques for further investigation.