Introduction
Stereo vision remains a critical capability for mobile robotics, providing the depth information essential for localization, mapping, and navigation. However, the computational overhead of traditional stereo matching often exceeds the capabilities of small-scale autonomous robots. The Keystone Scavenger Team addresses this challenge by introducing a "pragmatic" approach to stereo vision, optimized for hardware with limited processing power. By abstracting complex image regions into simplified geometric signatures, they achieve real-time performance on embedded systems.
Hardware and System Architecture
The research utilized two distinct robot platforms, "Jet" and "Slug," designed for Urban Search and Rescue (USAR) tasks. These robots are powered by 533 MHz C3 Eden VIA Mini-ITX boards—modest hardware by modern standards. To maximize efficiency, the team employed a custom miniaturized Debian Linux distribution with real-time extensions. This software architecture allows the main CPU to handle both high-level vision processing and low-level hardware control, such as PWM for servos, without the need for dedicated microcontrollers.
Methodology: Region-Based Stereo Vision
The core innovation of this work is a vision pipeline that prioritizes speed by avoiding pixel-by-pixel matching. The process involves six stages: color correction, blurring, edge detection, region extraction, simplification, and matching.
Convex Hull Signatures
To simplify the matching process, the team developed "convex hull signatures." After extracting image regions using a stack-based growing algorithm, they compute the convex hull of each region. A signature is then generated by measuring the distance from the region's center to the hull at 5-degree increments. This reduces a complex, high-dimensional shape into a simple array of 72 values.
Matching two regions is performed by calculating the sum of squared differences between these 72-value arrays. This geometric abstraction makes matching time independent of the region's actual pixel count, significantly reducing computational load. Furthermore, the system allows for overlapping segmentations, where the final region selection is determined by which candidate yields the highest-quality stereo match.
Key Results and Performance
The implementation of convex hull signatures provided a measurable performance boost during the 2006 AAAI Robotics competition:
- Frame Rate: The optimized system achieved a frame rate of 2.3–2.5 fps.
- Efficiency Gain: Without these geometric simplifications, the frame rate dropped to 1.5–1.8 fps.
- System Overhead: Given the hardware's base capture rate of 2.5–2.9 fps, the vision processing added minimal overhead, operating near the theoretical maximum of the hardware.
Beyond raw speed, the system successfully integrated Haar-feature-based face detection. By combining face recognition with stereo depth data, the robots could identify humans and navigate to a precise distance from them in unstructured environments.
Limitations and Future Work
While 2.4 fps is sufficient for the deliberate movement required in search and rescue, it remains a bottleneck for higher-speed maneuvers. Current human detection is also limited to facial recognition. Future work aims to expand detection capabilities to include other human features like hands and feet, which are more likely to be visible in disaster scenarios. Additionally, the team plans to integrate these region-based signatures into more robust visual SLAM (Simultaneous Localization and Mapping) frameworks.
Summary updated