When Dijkstra Meets Vanishing Point: A Stereo Vision Approach for Road Detection

Area of Science:

• Computer Vision
• Robotics
• Autonomous Driving

Background:

• Road detection is crucial for autonomous navigation.
• Existing methods often struggle with varying road conditions and complex scenes.
• Stereo vision offers depth information for improved road perception.

Purpose of the Study:

• To propose a novel, training-free road detection model using stereo vision.
• To leverage vanishing points and Dijkstra's algorithm for robust road boundary identification.
• To accurately extract road features like horizon and vanishing point.

Main Methods:

• Utilizing stereo-camera for u/v disparity map generation and horizon extraction.
• Robustly locating the road's vanishing point using horizon and ground constraints.
• Constructing a weighted graph with pixels as nodes, using disparity and gradient costs.
• Applying a vanishing-point constrained Dijkstra algorithm to find shortest paths for road detection.

Main Results:

• The model accurately detects horizons, vanishing points, and road regions.
• Achieved robust performance across 2600 diverse grayscale road images from the KITTI dataset.
• Demonstrated promising performance in challenging road detection scenarios.

Conclusions:

• The vanishing-point constrained Dijkstra road model offers a highly accurate and robust solution for road detection.
• This training-free approach simplifies implementation and broadens applicability in real-world autonomous systems.
• The method effectively integrates stereo vision cues for superior road perception.