Automatic Multi-Camera Extrinsic Parameter Calibration Based on Pedestrian Torsors.

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Traditional extrinsic camera calibration relies on physical objects, which is costly and impractical for field installations and recalibrations.
• Existing methods often require controlled environments and manual intervention, limiting their applicability.

Purpose of the Study:

• To propose a novel, accurate, and fully automatic extrinsic calibration framework for camera networks with partially overlapping views.
• To leverage pedestrians as natural calibration objects, reducing the need for artificial targets.

Main Methods:

• Detects human poses and models walking individuals as vertical sticks.
• Employs a brute-force approach for correspondence matching across multiple camera views.
• Utilizes 3D pedestrian locations (top and bottom points) to compute extrinsic calibration matrices.
• Introduces a method using only the top and centerline for occluded scenes.

Main Results:

• Achieves triangulation errors of a few centimeters.
• Requires less than one minute of observation in controlled environments for calibration.
• Demonstrates robustness in various scenarios, including multi-person, occlusions, and uncontrolled environments like street intersections.

Conclusions:

• The proposed pedestrian-based method offers a fully automatic, robust, and accurate solution for extrinsic camera network calibration.
• It significantly reduces the cost and complexity associated with traditional calibration techniques.
• The framework is adaptable to diverse and challenging real-world scenarios.