Modular Traffic Sign Recognition applied to on-vehicle real-time visual detection of American and European speed limit signs

We present a new modular traffic signs recognition system, successfully applied to both American and European speed limit signs. Our sign detection step is based only on shape-detection (rectangles or circles). This enables it to work on grayscale images, contrary to most European competitors, which eases robustness to illumination conditions (notably night operation). Speed sign candidates are classified (or rejected) by segmenting potential digits inside them (which is rather original and has several advantages), and then applying a neural digit recognition. The global detection rate is ~90% for both (standard) U.S. and E.U. speed signs, with a misclassification rate <1%, and no validated false alarm in >150 minutes of video. The system processes in real-time ~20 frames/s on a standard high-end laptop. INTRODUCTION AND RELATED WORKS Automatic traffic signs detection and recognition (TSR) is a key module for new driving assistance smart functions, as it is a requirement for the necessary level of traffic scene understanding. For example a robust visual real-time TSR system is a pre-requisite for developing a system for reminding the driver what is the current speed limit. Some of the traffic sign information may sometime be extracted from the GPS navigation data, but it is neither always complete nor up-to-date. Moreover, temporary speed-limits for road works, as well as variable speed-limits, are by definition not included in pre-defined digital cartographic data. Therefore a visual real-time TSR system is a mandatory complement to GPS systems for designing advanced driving assistance systems. A TSR system usually involves two main steps: 1/ detection of potential traffic signs in the image, based on the common shape/color design of sought traffic signs; 2/ classification of the selected regions of interest (ROI) for identifying the exact type of sign, or rejecting the ROI. As noted by Bahlman et al. in [1], the majority of recently published TSR approaches make use of color information (see e.g. [2], [3] or [4]), which makes the detection step easier. In contrast with that, the TSR system presented in the present paper works on grayscale images, which puts less constraint on the required sensor, and may help meet global costs