Multi-Camera 3D Human Pose Reconstruction System

Reconstructs accurate 3D human pose from multiple synchronized camera views, for sports or physiotherapy motion analysis.

A single camera sees a person in two dimensions, so depth is lost and limbs that overlap cannot be measured. Sports coaches and physiotherapists need real 3D joint positions to judge technique or recovery, and specialised motion-capture suits are expensive. This project reconstructs accurate 3D human pose from several ordinary synchronized cameras.

The system first calibrates each camera, estimating its lens distortion and its position relative to the others using a checkerboard or a reference object. It then runs a 2D pose estimator on every video frame from every camera to find joints such as knees and elbows. Using the calibration, the matching 2D joints from different views are triangulated into 3D points, with outlier rejection and temporal smoothing to remove jitter. Open3D renders the skeleton in a 3D viewer, and a motion-analysis tool computes joint angles, range of motion and movement speed so that a coach can compare repetitions. Accuracy is evaluated against known reference measurements and the effect of the number of cameras is studied.

You will learn camera geometry, calibration, multi-view triangulation and how to evaluate 3D accuracy. The Project Reference Guide is written at research depth, and the Reference Implementation includes the calibration tools, reconstruction code, viewer and set-up instructions.