Accurate Measurement of 3D and 2D Circular Centers With Application to LiDAR-Camera Extrinsic Calibration

In the world of machines that see, precision is everything. For a robot to navigate a room, a self-driving car to avoid a pedestrian, or a drone to land safely, it must know exactly where it is in relation to the objects around it. This task often relies on two different kinds of eyes: a camera, which captures a flat, two-dimensional picture of the world, and a LiDAR sensor, which shoots out invisible laser beams to map the three-dimensional shape of objects in space. To make these two sensors work together as a single team, engineers must perform a delicate alignment process called extrinsic calibration. This is the act of figuring out exactly how the camera and the LiDAR are positioned relative to one another. If this alignment is off by even a tiny fraction, the robot's understanding of the world becomes distorted, leading to errors in navigation or measurement. A common way to perform this alignment is to use a simple, circular target, like a painted disk or a metal ring, and ask both sensors to find its center. The challenge, however, is that finding the true center of a circle is surprisingly difficult when the data is noisy, incomplete, or viewed from an angle.

A team of researchers has tackled this specific problem, focusing on the hidden errors that occur when measuring the centers of circles in both 3D space and 2D images. In standard practice, when a LiDAR sensor scans a circular object, the data it collects is often a scattered cloud of points. The traditional method for finding the center of this shape involves a two-step process: first, the computer fits a flat plane through the scattered points, and then it projects those points onto that plane to fit a simple 2D circle. The researchers found that this step-by-step approach introduces a systematic bias, meaning the calculated center is consistently shifted away from the true center, especially when the data is sparse or the circle is only partially visible. Similarly, when a camera looks at a 3D circle, the circle rarely appears as a perfect circle in the image; perspective distortion usually turns it into an oval, or ellipse. Conventional methods often assume that the center of this oval is the same as the center of the original 3D circle. The researchers demonstrated that this assumption is incorrect; under most viewing angles, the center of the oval is actually displaced from the true projection of the 3D circle's center. These small errors in the initial measurements accumulate, causing the final alignment between the camera and the LiDAR to be less accurate than it could be.

To solve this, the researchers developed a new approach that treats the measurement of the 3D circle and the 2D image as a unified, geometric problem rather than a series of separate steps. For the 3D data from the LiDAR, they introduced a method that uses a specialized mathematical framework to fit the circle directly to the points in space. Instead of flattening the data first, this method finds the best-fitting sphere and plane simultaneously, effectively ignoring the noise and outliers that usually throw off traditional calculations. This technique is robust enough to handle situations where the sensor only sees a small arc of the circle or where the data is contaminated by random errors. For the 2D image, the team created a new way to locate the true center of the projected circle. They realized that if you draw lines through the true center of the projected shape, the lengths of the chords they create on the oval follow a specific, predictable pattern. By searching for the point where the variation in these chord lengths is minimized, the system can pinpoint the true center, distinguishing it from the misleading center of the oval shape.

A significant hurdle in this 2D measurement is that the mathematical search for the true center often produces two possible answers, or "local minima," and it is not immediately obvious which one is correct. The researchers solved this ambiguity in two ways. If the calibration target contains two circles lying on the same flat surface, the system can use the geometric relationship between them to verify which center is the right one. If only a single circle is available, they devised a fallback strategy that tests both possibilities against the 3D data to see which one results in a consistent alignment. The team tested these methods using computer simulations, synthetic data, and real-world experiments with various sensors, including solid-state and mechanical LiDARs paired with different types of cameras. The results showed that their new methods significantly reduced the error in measuring the circle centers. In simulations, the new approach reduced the error in the final alignment between the camera and LiDAR from over a centimeter to less than a centimeter, and it improved the precision of the rotation angle as well.

The study confirms that the accuracy of a robot's perception depends heavily on the quality of the geometric primitives it uses to understand the world. By correcting the systematic biases in how circular centers are measured, the researchers have provided a more reliable foundation for calibrating heterogeneous sensor systems. Their work does not just offer a theoretical improvement; it provides a practical tool that works across different sensor configurations and real-world conditions, from industrial manufacturing to autonomous driving. The researchers have made their code publicly available, allowing other engineers to apply these more accurate measurement techniques to their own systems. While the study focused on circular targets, the underlying principle—that direct, joint optimization of geometric measurements is superior to sequential, decoupled steps—suggests a path forward for improving the accuracy of many other types of sensor fusion tasks. The findings underscore that even the simplest geometric shapes require sophisticated handling when viewed through the imperfect lenses of real-world sensors.