Abstract
Visual surveillance using camera networks presents significant challenges in calibration, particularly when cameras lack a common field of view or precise synchronization. This paper proposes a hybrid camera network architecture combining catadioptric (omnidirectional) and perspective cameras. By utilizing the catadioptric camera as a global coordinate anchor, the system enables unified calibration across diverse views. We present a three-step calibration framework: first, calibrating the catadioptric system using vanishing points; second, localizing 3D points via parallel line geometry; and third, mapping perspective cameras to the global system. Experimental results demonstrate 3D localization errors of less than 3.5%, proving the effectiveness of this hybrid approach.
The Hybrid Camera Network Challenge
In visual surveillance, practitioners often face a fundamental trade-off. Standard perspective cameras provide high-resolution details necessary for identification but suffer from narrow fields of view (FOV) and frequent occlusions. Conversely, omnidirectional (catadioptric) cameras provide a 360-degree situational awareness but at a significantly lower resolution.
The primary technical hurdle is establishing a unified global coordinate system. Traditional calibration methods require overlapping fields of view to observe shared calibration patterns. In large-scale deployments, many cameras do not overlap, and synchronization is rarely perfect. This research addresses these gaps by using the wide-angle catadioptric camera as a "global root" for the entire network.
Methodology
The proposed framework follows a top-down calibration procedure in three distinct stages:
Step 1: Catadioptric Calibration
The first step involves calibrating the omnidirectional camera. To manage complexity, the researchers first estimate the camera's aspect ratio without the mirror. They then leverage the geometry of vanishing points. In a parabolic catadioptric system, parallel lines in the physical world project as circular arcs on the image plane. By fitting these arcs to a circle, the system can determine the principal point and the combined focal length.
Step 2: 3D Point Localization
Once the catadioptric camera is calibrated, it serves as the origin of the coordinate system. The authors utilize geometric propositions to determine the 3D positions of specific environmental markers. This is achieved using as few as two sets of parallel lines and known equidistance points, allowing the environment itself to act as a calibration rig.
Step 3: Perspective Camera Calibration
Finally, standard perspective cameras are integrated into the network. As long as a perspective camera can view at least six of the 3D points localized in the previous step, its intrinsic and extrinsic parameters can be mapped back to the global coordinate system defined by the omni-camera. This "anchor" strategy simplifies the network topology from a complex graph of overlaps into a manageable star pattern.
Experimental Evaluation
The method was evaluated using a custom laboratory setup consisting of one Cyclovision omnidirectional camera and four Sony perspective cameras. Instead of specialized checkerboard patterns, the researchers used structural features of the room, such as ceiling lines, to establish parallel line constraints.
Key Results: * High Precision: The system estimated 3D coordinates with a consistent error rate of less than 3.5% compared to manual ground-truth measurements. * Robust Focal Estimation: The algorithm successfully calculated the complex focal length (113.13 pixels) of the catadioptric system. * Seamless Hand-offs: Re-projection of 3D points onto perspective views showed minimal pixel error, indicating that the network can accurately hand off tracking data from a wide-angle overview to high-resolution detail views.
Conclusion and Future Directions
This hybrid calibration framework effectively bridges the gap between different camera types in a network. While the system shows high accuracy, performance can degrade near the edges of the catadioptric mirror due to non-linear mapping and lower resolution. Future work aims to automate the selection of parallel lines and markers, moving toward a fully self-calibrating autonomous surveillance architecture.
Summary updated