Intelligent road sign detection using 3D scene geometry
Jeffrey Schlosser, Michael Montemerlo, Ken Salisbury · 2010
This paper proposes a new framework for fast and reliable traffic sign detection using images obtained from a single front-facing road vehicle camera. Our focus is on a methodology for reducing the computational requirements and increasing the performance of existing detection methods by refining the image space search using 3D scene geometry. Information concerning physical traffic sign dimensions and vehicle camera parameters is integrated into a model that predicts the image scales and locations at which traffic signs are likely to appear. We apply our framework to a Haar-feature-based detection method trained on a collection of stop signs. Experimental results show that the refined image search space results in much less computation time while retaining the same true positive detection performance as existing methods that search all image scales and locations. In addition, false positives at physically implausible traffic sign locations are eliminated.