Sheep pose estimation via image analysis and body measurements derived from key points

Pose estimation and body measurement in livestock play an important role in various agricultural applications such as health monitoring, breeding and management. In this study, we propose a novel approach for body measurement and pose estimation of sheep using object detection and supervised machine learning algorithms. From a dataset of sheep, 40 side-view videos for each class (grazing, standing, and sitting), each lasting approximately 8 seconds, are used for model training. Firstly, pose classification was performed on the sheep images using the YOLOv8 pose object detection framework. Simultaneously, keypoint training was performed on the images to detect key points for anatomical landmarks. Then, using these keypoints, various body measurements of the sheep in the images were measured and a comprehensive dataset was created. Six different supervised machine learning algorithms were trained on this dataset to further improve pose estimation. Furthermore, the models were tested on frontal images to evaluate their performance against different image angles and dataset feature. The experimental results show that the supervised machine learning algorithms trained on the body measurement data perform better for both side and frontal images (mAP; K nearest neighbour algorithm 1.00 for side images, Support vector machines 0.94 for frontal images) outperform state-of-the-art networks such as YOLOv8 (mAP; 0.94, 0.89 for side and frontal images respectively), EfficientNet (mAP; 0.93, 0.91), RetinaNet (mAP; 0.91, 0.88) and Faster R-CNN (mAP; 0.89, 0.87) based on image data only. This approach can play an important role in improving the accuracy and efficiency of livestock management systems by supporting practical applications such as animal welfare monitoring, herd health assessment, and precision agriculture through more accurate position estimation and body measurements.