Introduction
Omni-directional (fish-eye) cameras are a staple in surveillance and robotics due to their expansive 180-degree field of view. However, this breadth comes at the cost of extreme geometric distortion. In computer vision, Zernike moments are widely used to describe object shapes because they are mathematically designed to be invariant to rotation and scale. However, these moments assume a flat, Euclidean image plane. In a fish-eye image, an object appears "stretched" at the edges compared to the center, causing traditional Zernike moments to lose their descriptive power and leading to poor classification performance.
Methodology: Geodesic Correction
The authors propose a Geodesic Correction for calculating Zernike moments. Instead of treating the image as a flat grid, they treat it as a projection of a sphere. By performing calculations on the sphere, the method effectively "undoes" lens distortion without the high computational cost of unwarping the entire image into a panoramic view.
- Spherical Mapping: Using a calibrated camera model, every pixel $(i, j)$ on the image sensor is mapped to coordinates $(\theta, \phi)$ on a virtual unit sphere.
- Geodesic Distance ($r$): Instead of straight-line distance on a flat sensor, the method calculates the arc length (the shortest path along the sphere's surface) between a pixel and the object's centroid.
- Geodesic Angle ($\gamma$): The angle is redefined by calculating the angle between the tangent vectors of two circles on the sphere: the "great circle" passing through the pixel and the centroid, and the "parallel circle" passing through the centroid.
To maintain real-time efficiency, the algorithm uses finite differences to approximate tangent vectors, avoiding heavy trigonometric computations. Additionally, using ratios of different moment orders (e.g., $Z_{40}/Z_{20}$) provides scale invariance, making the system robust as subjects move toward or away from the camera.
Experimental Results
The researchers evaluated the method using geometric control sequences and human pose datasets (walking/standing vs. sitting).
- Improved Accuracy: In human pose classification using a $k$-Nearest Neighbor ($k=3$) classifier, the geodesic correction consistently outperformed traditional methods. Accuracy for radial Zernike moments improved from 94.53% to 96.13%.
- Data Stability: Qualitative analysis shows that geodesic Zernike moments are significantly less noisy than traditional ones, providing a consistent "signature" for a shape regardless of its position in the high-distortion areas at the lens edges.
- Real-Time Performance: Despite the spherical mathematics, the system processed between 9 and 17 frames per second (FPS), making it suitable for live monitoring.
Conclusion
Geodesic correction allows for robust shape description in omni-directional vision by accounting for the underlying spherical geometry of the lens. This methodology provides a foundation for recognizing complex activities in Ambient Assisted Living (AAL) environments, such as fall detection, where wide-angle coverage is essential.
Summary updated