Background This continuation advances our previous NAISS projects that couples AI-based, monocular markerless motion tracking with finite-element (FE) human body models (HBMs) to analyze injury mechanisms from real-world traffic and sports video. It builds directly on earlier projects that established the data layout, containerized toolchain, and the first monocular video→FE coupling on Alvis. Motivation and research gaps Despite strong progress in computer vision and HBMs, three gaps remain for in-the-wild, single-view footage: (1) Estimator sensitivity: FE injury metrics (e.g., head kinematics, neck loads) can be highly sensitive to upstream choices in SMPL/SMPL-X estimation and hyper-parameters, yet this sensitivity is under-quantified for unconstrained video. (2) Domain shift: Fast motion, occlusions, camera shake, and broadcast compression common in traffic and field/ice sports degrade pose stability and tracking continuity. (3) Reproducibility at scale: A robust, containerized video→SMPL→FE pipeline—auditable end-to-end and scalable on cluster GPUs—has not been broadly demonstrated on diverse real-world incidents. Objectives We aim to (i) quantify how SMPL-family estimator variants and configurations propagate to FE-level injury metrics; (ii) adapt multi-person pose/tracking to fast-motion traffic and sports domains via light fine-tuning and curated data; and (iii) scale a reproducible monocular video→FE workflow to large-batch inference on Alvis. Key questions include: How large are estimator-induced variations in FE outcomes? Which model choices most affect head/neck responses? What adaptation strategies most improve stability on fast motion and compression? Approach Containerized PyTorch stacks (Apptainer/Singularity) execute a logged lineage: decode → detect/track → SMPL/SMPL-X fit → export FE-ready 3D kinematics (head/torso focus). We perform sensitivity sweeps across estimator families and hyper-parameters, propagate trajectories to HBMs, and analyze injury metrics (e.g., peak angular velocity/acceleration, strain-based criteria). Data domains span traffic cameras/dashcams and broadcast sports (e.g., ice hockey, American football), emphasizing monocular constraints. All directories, metadata, and job templates adhere to the previously approved Alvis structure for reproducibility. Validation Tracking stability and keypoint quality are evaluated on held-out clips with failure-mode auditing (fast motion/occlusion). FE predictions are compared against documented incident characteristics and available medical evidence. We report both algorithmic metrics (throughput, drop rate) and biomechanical metrics (variance and bias in FE responses under estimator changes). Expected outcome. An adapted, generalizable multi-person tracking/pose stack robust to fast motion and compression; A reproducible, containerized pipeline suitable for large scale batch processing, with non-identifiable derived kinematics for sharing;