Contacts: Marta Gandolla, Andrea Dal Prete
Location: Politecnico di Milano (Bovisa Campus)
Motivation of the study: Accurate human kinematics measurement is fundamental in biomechanics, ergonomics, and human–robot interaction, yet traditional approaches remain limited. Optical motion capture systems deliver high accuracy but require expensive equipment and controlled laboratory conditions, while IMU-based methods offer portability at the cost of drift and variability. These constraints motivate the search for scalable and non-intrusive alternatives.
Recent advances in AI-driven pose estimation and 3D body reconstruction, such as emerging foundation models (e.g., SAM3D Body) and lightweight pipelines like MediaPipe, enable full-body mesh reconstruction directly from video, potentially providing rich kinematic information without specialised hardware. Despite their promise, their reliability and accuracy relative to established measurement systems remain insufficiently understood.
This thesis aims to benchmark modern AI-based pose estimation and 3D reconstruction techniques against traditional motion capture and IMU-based methods. By evaluating their performance across different conditions and kinematic demands, the study seeks to determine whether these innovative AI approaches can serve as practical and accessible substitutes for conventional human motion measurement techniques.
Objective: The candidate will conduct a systematic benchmarking study of innovative AI-driven pose estimation and 3D human body reconstruction techniques for kinematic analysis. The work involves reviewing relevant literature, implementing and evaluating selected AI methods, including foundation-model-based 3D reconstruction and lightweight pose estimation pipelines like SAM3D Body (blog, video), and comparing them against traditional motion capture and IMU-based approaches. The candidate will analyze accuracy, robustness, and practical usability across different movements and conditions, and integrate the results into a comprehensive performance assessment. The ultimate goal is to determine whether modern AI-based strategies can serve as reliable and accessible alternatives for human kinematics measurement, informing future research and applications in biomechanics and human-motion analysis.
Credit: Yang, Xitong and Kukreja, Devansh and Pinkus, Don and Sagar, Anushka and Fan, Taosha and Park, Jinhyung and Shin, Soyong and Cao, Jinkun and Liu, Jiawei and Ugrinovic, Nicolas and Feiszli, Matt and Malik, Jitendra and Dollar, Piotr and Kitani, Kris. SAM 3D Body: Robust Full-Body Human Mesh Recovery, arXiv preprint, 2025.
