Abstract
To enhance the real-time performance of indoor mobile robot visual positioning, this research investigates how image shape and size influence positioning calculations. Using the SURF algorithm for feature extraction, the study evaluates four geometric shapes (quadrilateral, circle, triangle, and hexagon) across varying image scales. Results indicate that squares and circles provide superior accuracy and speed compared to triangles and hexagons. Furthermore, reducing image size to 40–60% of the original dimensions optimizes the balance between feature point stability and computational efficiency, significantly improving real-time localization without compromising precision.
Introduction
Indoor mobile robots rely on vision systems for navigation and environmental interaction. However, high-resolution image processing often introduces significant computational latency. While algorithmic optimizations for SIFT or SURF are common, the impact of environmental target properties—specifically shape and input resolution—remains less explored. This paper aims to identify the optimal "sweet spot" for image parameters to ensure responsive, real-time robot positioning.
Methodology
The researchers employed the Speeded-Up Robust Features (SURF) algorithm, favored in robotics for its speed, utilizing integral images and Hessian matrix approximations for rapid interest point detection.
As illustrated in Figure 1, the experimental workflow consists of several key stages:
- Image Preprocessing: Conversion of color images to grayscale and iterative threshold segmentation to isolate targets from the background.
- Feature Extraction: Identifying keypoints via the SURF algorithm.
- Shape Benchmarking: Testing quadrilateral (square), circular, triangular, and hexagonal templates to determine which provides the most stable feature points.
- Size Benchmarking: Systematically scaling images from 10% to 100% of an 800x600 baseline to measure the impact on computational load and point density.
Experimental Setup and Datasets
Experiments were conducted using a four-axis vision guidance platform and an Epson robotic arm. A custom dataset was generated featuring the four geometric templates under controlled indoor lighting. The classic "Lena" image served as a baseline for validating the effectiveness of the grayscale and segmentation preprocessing steps.
Key Insights
Symmetry and Stability
Highly symmetrical shapes, such as squares and circles, yield more consistent feature points for the SURF algorithm. This symmetry reduces tracking errors and accelerates computation compared to more complex or "pointy" shapes like triangles and hexagons.
The Resolution Plateau
The study found that increasing image size beyond a certain threshold offers diminishing returns. While higher resolution provides more pixels, the number of useful feature points eventually stabilizes. Processing redundant pixels increases latency without a corresponding gain in positioning accuracy.
The "Goldilocks" Zone for Scaling
A critical finding is the existence of an optimal scaling window. At 40% to 60% of the original image size, the number of feature points is sufficient for high accuracy while keeping the computational load manageable for real-time applications.
Results
- Positioning Accuracy: Circles and squares achieved the highest precision, with average positioning errors of 0.8 mm and 0.86 mm, respectively. In contrast, hexagons performed the worst with an error of 1.27 mm.
- Computational Speed: In 10-cycle positioning tests, squares (8850 ms) and circles (8860 ms) were significantly faster than triangles (9700 ms).
- Size Optimization: Scaling from 10% to 40% shows rapid growth in feature points, which then stabilizes between 40% and 100%. Operating in the 40–60% range minimizes calculation time variance, enhancing system reliability and real-time performance.
Conclusion and Future Work
Optimizing target shape and input resolution significantly boosts the real-time capabilities of robot vision systems. While this study provides a clear framework for indoor environments using SURF, future research should explore these parameters in dynamic, "messy" real-world settings and evaluate their applicability to newer, deep-learning-based feature detectors like SuperPoint or ORB.
Summary updated