Abstract
This paper introduces a computationally efficient "eye-in-the-sky" localization system for indoor mobile robots. Utilizing a single stationary fisheye camera mounted on the ceiling, the system maps 2D image coordinates to 3D world coordinates through a specialized geometric model and pre-calculated lookup tables. Experimental results demonstrate that this infrastructure-centric approach achieves sub-10cm accuracy in real-time without the need for complex onboard SLAM or power-intensive sensors.
Introduction
Localization is a fundamental requirement for autonomous robotics. Traditional methods typically rely on onboard sensing (e.g., LiDAR, sonar) or dead-reckoning via wheel odometry. While effective, onboard sensing is often computationally expensive, and odometry suffers from cumulative drift over time. This research proposes an alternative: moving the localization intelligence to the environment's infrastructure. By using a wide-angle fisheye camera, the system monitors the entire workspace, providing absolute positioning that eliminates incremental error.
Methodology
The proposed system operates in two distinct phases: a one-time calibration and a real-time tracking loop.
System Architecture
The workflow separates heavy computations from live operations. During calibration, the system calculates azimuth and elevation for every pixel based on a fisheye model and stores them in Lookup Tables (LUTs). This allows the real-time phase to perform coordinate transformations using simple table lookups rather than complex trigonometry.
Geometric Projection Model
To handle the significant "barrel" distortion inherent in fisheye lenses, the authors implemented a spherical image formation model. This model maps 3D world coordinates $(x, y, z)$ to 2D image pixels $(i, j)$. By deriving the inverse of this model, the system can determine the specific ray (defined by azimuth $\theta$ and elevation $\phi$) originating from the camera to the robot.
Tracking and Segmentation
The robot is identified via a color marker on its top surface. The system employs RGB-based color segmentation to isolate the marker. Since the robot's height ($h_R$) is constant and it moves on a flat floor, the $(x, y)$ ground coordinates are calculated by intersecting the camera's line-of-sight with the floor plane. To optimize performance, a search mask restricts processing to accessible floor areas, reducing false positives.
Experimental Setup
The system was validated using a PeopleBot (Adept MobileRobots) in an indoor environment. Ten video sequences (90–120 seconds each) were recorded at 640x480 resolution and 25 fps. Ground truth data was obtained from the robot’s internal sensors (odometry and onboard sensors) for comparison against the camera-based localization.
Results and Discussion
The system demonstrated high precision and efficiency:
- Accuracy: Within a 4.5-meter radius of the camera, the positional error remained under 0.1 meters (10 cm).
- Temporal Performance: The algorithm processed frames in approximately 70ms on standard consumer hardware. Real-world throughput reached 7 fps over WiFi, which is sufficient for typical indoor robot velocities.
- Error Characteristics: Unlike odometry, the error is absolute and does not accumulate. However, accuracy is non-linear; spatial resolution is highest directly beneath the camera and degrades toward the periphery of the fisheye lens.
Conclusion
This "eye-in-the-sky" approach provides a robust, low-cost solution for indoor localization. While limited by line-of-sight requirements and decreasing resolution at long distances, the system offers a drift-free alternative to traditional onboard methods. Future work will focus on porting the implementation from Matlab to C++ to enhance frame rates and exploring evolutionary algorithms for automated camera calibration.
Summary updated