MARMOT
Published:
Overview
Multi-camera localisation, calibration, and ground-plane pedestrian tracking system with automated axis-alignment. (1 year)
Problem
Deploying a multi-view tracking system requires robust camera localisation and calibration that can recover reliable world-frame geometry — including consistent axis orientation — from diverse camera setups with minimal manual intervention.
Approach
- Camera localisation and calibration: Led development of a multi-camera localisation and calibration suite for the multi-view tracking system
- Automated axis-alignment: Developed an automated axis-alignment method for OpenSFM reconstructions using RANSAC plane fitting on keypoint-dense regions combined with camera height priors to recover z-axis orientation
- Ground-plane tracking: Adapted ByteTrack and YOLOX for ground-plane operation by projecting detections onto the floor to enable meaningful cross-view correspondence
Technical Details
The system handles the full pipeline from raw multi-camera video to calibrated camera models and ground-plane pedestrian trajectories. The RANSAC-based axis alignment removes the need for manual ground-plane annotation, and the projected-detection approach enables robust cross-view association for tracking.