Road detection for safe autonomous navigation is a critical computer vision task focused on segmenting navigable areas from road image frames. Edge detection identifies object boundaries within an image and is foundational to tasks like segmentation, enhancement, and object tracking. While several edge detectors exist, this study evaluates two prominent edge detectors, Sobel and Canny operators, for drivable region detection. Both operators are evaluated on a stream of real-world road images across varying resolutions under clear and occluded scenarios. Drivable regions were detected using a binary edge mask and by identifying the largest connected component and performing a seeded flood fill in the road image with an adaptive morphological closing operation. Each edge detector’s performance with each drivable region extraction method was evaluated using Pixel accuracy (PA) and Intersection over Union (IoU). Experimental results on 80 images across six scene categories demonstrate that both detectors can identify drivable regions, but the region extraction method influenced accuracy more than the choice of edge detector. Seeded Flood Fill (FF) raised mean Pixel Accuracy for both detectors, and with FF, Sobel and Canny achieved near-identical mean IoU. However, FF was less consistent, scoring below 1% IoU on 20 to 21 images, mostly portrait-oriented ones. Sobel with FF was the fastest configuration, averaging 0.24 s per image, giving the best balance of accuracy and speed in this research.
Honours student
4262103@myuwc.za
Supervisor
oisafiade@uwc.ac.za
Co-Supervisor
amaneli@uwc.ac.za