А. В. Греченева Body Condition Score (BCS) in dairy cattle is used to monitor animal energy status and to identify the risks of metabolic and reproductive disorders. In industrial video recording, the accuracy of automated BCS estimation is limited by the displacement of reference anatomical landmarks. This displacement arises due to changes in animal posture, walking speed, uneven lighting, glare, and shadows. The study aimed to develop an adaptive video-frame selection algorithm to improve the detection accuracy of reference keypoints and contour landmarks in automated body condition assessment of dairy cattle. The research material consisted of video sequences of 983 dairy cows. For analysis, a system of reference points and contour landmarks of the pelvic region was used. Image processing was performed using a multitask neural network model for localizing key points and contours, followed by BCS prediction on a 1-5 scale. The developed algorithm included adaptive frame sampling based on animal walking speed, visibility assessment of anatomical zones, and component-wise selection of the most informative frames to form a consistent set of features. Application of the algorithm resulted in a reduction of the mean absolute error in BCS estimation from 0.34 to 0.22 score points, an increase in the proportion of predictions within ±0.5 points from 86.2% to 93.7%, and an improvement in the weighted Cohen’s κ coefficient from 0.74 to 0.86. The normalized localization error of reference keypoints decreased from 0.071 to 0.048. The findings confirm that adaptive frame preprocessing enhances the accuracy and robustness of automated body condition assessment under industrial video recording conditions.
Научная статья 
нейросетевая модель