Markerless 3D biomechanical analysis of manual rice transplanting using YOLO11 and depth anything V2

Manual rice transplanting operation imposes severe biomechanical demands on agricultural workers, but quantitative assessment has been limited by practical limitations of marker-based motion capture in agricultural fields. This study presents a validated markerless framework integrating YOLOv11 pose estimation with Depth Anything V2 monocular depth prediction that is validated and field-deployable for three-dimensional biomechanical analysis. This framework extracts 17 anatomical keypoints and depth maps from standard video recordings (1080p, 30 fps). Full-body kinematics with 36 degrees of freedom are reconstructed following International Society of Biomechanics conventions. Numerical differentiation was performed for determining angular velocities, accelerations, and jerk along with Butterworth filtering (6 Hz cutoff) to reduce noise amplification inherent in high-order derivatives. Manual transplanting analysis (1366 frames, 48 transplanting cycles, 45.8 s, 64 hills/minute) yielded 25,276 validated kinematic measurements revealing three critical hazards: (1) Sustained trunk flexion (89.8 ± 3.3°) exceeded the 60° ergonomic threshold by 49 %, corresponding to 3.5–4.5 times bodyweight spinal compression; (2) Shoulder accelerations surpassed 600°/s² safety limits in 83.7 % of frames (peak: 11,603°/s²), indicating elevated rotator cuff injury risk; (3) High jerk values in 48.3 % of movements (mean: 1545°/s³) signaled poor motor control and tissue strain. The kinematic dataset provides robotic design specifications: peak joint velocities (350–400°/s), accelerations (450–700°/s²), and end-effector speeds (∼2.8 m/s). This validated markerless approach enables scalable field-based biomechanical assessment, reducing infrastructure costs by >99 % ($200K-500 K to $1K-1.5 K) while maintaining measurement rigor for occupational health evaluation and robotic system development.