YOLOv8-DMC: Enabling Non-Contact 3D Cattle Body Measurement via Enhanced Keypoint Detection.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Animal Science

Background:

• Precision livestock management requires accurate, non-contact methods for measuring cattle body dimensions.
• Existing methods may lack accuracy or robustness in real-world farming environments.

Purpose of the Study:

• To develop and validate YOLOv8-DMC, a lightweight deep learning model for precise 3D cattle measurement using side-view images.
• To improve the accuracy and robustness of anatomical keypoint detection in cattle under challenging conditions.

Main Methods:

• Utilized YOLOv8 architecture integrated with DRAMiTransformer, MHSA-C2f, and CASimAM attention modules for keypoint detection.
• Implemented a 16-neighborhood depth completion and filtering process for generating clean, colored point clouds.
• Validated the model on a dataset of over 7000 images and real-world RGB-D images from 137 cattle.

Main Results:

• Achieved high accuracy with AP@0.5 of 0.931 and AP@[0.50:0.95] of 0.868.
• Demonstrated improved accuracy over baseline by 2.14% and 3.09% with minimal increase in parameters and complexity.
• Reported low average relative errors for body height (2.43%), hip height (2.26%), body length (3.65%), and cannon circumference (4.48%) compared to manual measurements.

Conclusions:

• YOLOv8-DMC offers an efficient and accurate solution for 3D cattle measurement.
• The model's robustness to occlusion and lighting variability makes it suitable for real-world farming.
• The system's support for edge device deployment facilitates practical application in precision livestock management.