Automatic estimation of dairy cow body condition score based on attention-guided 3D point cloud feature extraction

Abstract

Body condition score (BCS) is an important indication in management of dairy cow breeding, which can be used to evaluate milk production, reproduction and health status of dairy cows. Due to the low efficiency and subjectiveness of manual scoring, it’s necessary to estimate BCS of dairy cows automatically and accurately. The current automatic estimation methods generally utilized features extracted from 2D image data or manually defined 3D surface features of dairy cows’ back end, which were defective or difficult to essentially represent the 3D concavity information of local area of dairy cattle body severed as the key indicator of BCS. In this study, a 3D data format dataset was first built for dairy cow body condition score estimation. To learn a more effective representation of concavity information automatically and focus on vision saliency information precisely of local area of dairy cows’ back end, an automatic method was proposed to estimate BCS of dairy cows based on attention-guided 3D point cloud feature extraction. Experiments show that proposed body condition scoring model achieved the accuracy of 0.49, 0.80, 0.96 within 0, 0.25, 0.50 point deviation respectively, which has achieved good estimation results in comparison with the other research.