Traffic sign recognition using MSER and Random Forests

We present a novel system for the real-time detection and recognition of traffic symbols. Candidate regions are detected as Maximally Stable Extremal Regions (MSER) from which Histogram of Oriented Gradients (HOG) features are derived, and recognition is then performed using Random Forests. The training data comprises a set of synthetically generated images, created by applying randomised distortions to graphical template images taken from an on-line database. This approach eliminates the need for real training images and makes it easy to include all possible signs. Our proposed method can operate under a range of weather conditions at an average speed of 20 fps and is accurate even at high vehicle speeds. Comprehensive comparative results are provided to illustrate the performance of the system.