Accurate and non-contact measurement of cattle body dimensions is essential for precision livestock management. This study presents YOLOv8-DMC, a lightweight deep learning model optimized for anatomical keypoint detection in side-view images of cattle. The model integrates three attention modules—DRAMiTransformer, MHSA-C2f, and CASimAM—to improve robustness under occlusion and lighting variability. Following keypoint prediction, a 16-neighborhood depth completion and pass-through filtering process are applied to generate clean, colored point clouds. This enables precise 3D localization of keypoints by matching them to valid depth values. The model achieves $A{P}_{@0.5}$ of 0.931 and AP@[0.50:0.95] of 0.868 on a dataset of over 7000 images, improving baseline accuracy by 2.14% and 3.09%, respectively, with only 0.35 M additional parameters and 0.9 GFLOPs in complexity. For real-world validation, strictly lateral-view RGB-D images from 137 cattle were collected, with ground-truth manual measurements. Compared with manual measurements, the average relative errors are 2.43% for body height, 2.26% for hip height, 3.65% for body length, and 4.48% for cannon circumference. The system supports deployment on edge devices, providing an efficient and accurate solution for 3D cattle measurement in real-world farming conditions.