Automatic estimation of dairy cattle body condition score from depth image using ensemble model

Body condition scoring (BCS) gives a relative measure of subcutaneous body fat available as energy reserves in the dairy cow. It is an important management tool for maximising milk production and reproduction efficiency while reducing the incidence of metabolic and peripartum diseases. The feasibility of estimating the BCS by computer vision has been demonstrated in recent research. However, the techniques explored to date may be limited in dynamic backgrounds or in applications for an imbalanced dataset of cows' BCS, which is likely to be encountered in dairy farming. In this study, a dynamic background model (Gaussian Mixture Model, GMM) was used to separate the cow from the background. Then, a series of image processing algorithms were proposed for quantifying the indicators used in manual scoring, including global features and local features. Finally, an ensemble learning approach was used to model the imbalanced dataset. The results demonstrate that applying GMM on depth images can eliminate the difficulty of object detection caused by background changes. The image processing algorithms can automatically acquire valid images, locate regions of interest and extract image features without any manual intervention. In 5-fold cross-validation, the ensemble model achieved an average accuracy of 56% within 0.125-point deviation, 76% within 0.25-point deviations and 94% within 0.5-point deviations. Especially, the proposed method has a better predictive performance for cows with extreme body condition than is possible with the current state of the art.