Traffic Sign Recognition Using Sparse Representations and Active Contour Models

Abstract

Automated traffic sign recognition (TSR) remains an integral component of road sign inventory programs administered by transportation agencies. However, current practices for TSR are labor-intensive and time-consuming. This paper presents a novel TSR algorithm that can be integrated into the traffic sign management systems of infrastructure managers. The methodology proposed the use of a multiscale principal component pursuit (mPCP) for detection of the locations of traffic signs that decomposes each video frame into a background image (trees, cars, illumination, etc.) and foreground image (traffic sign information). Next, a novel model method for edge detection called “active contours,” or “snakes,” was introduced to extract traffic sign geometries for classification. The active contour models showed promise for the detection of different types of traffic signs in noisy environments. The ability of these models to split and match topologies of the image data was essential for accurate sign location and shape detection. The evaluation of the proposed methodology was conducted with data from video logs. The performance of the system developed was compared with that of the Canny- and Hough-based TSR algorithms. The test results indicated that the mPCP–active contour-based TSR algorithm remarkably outperformed the selected benchmark algorithms, with the highest detection rate (92.6%) and precision (87.2%). These results demonstrated the potential for practical application of the proposed methodology.