A Hybrid Approach Of Candidate Region Extraction For Robust Traffic Light Recognition

In this paper, we consider the problem of recognizing circular traffic lights from an image. The traffic light recognition is divided into two stages: candidate region extraction and traffic light recognition. For extracting candidates, we propose a hybrid method, which combines the results of spotlight detection and color-shape model-based method. Instead of handcrafting a set of features for classification, we scale each candidate as a 10-by-10 image patch and use its raw RGB pixel values as the input of a Support Vector Machine (SVM) classifier. The classifiers are trained only using a Singapore dataset, and are tested on the US LISA dataset. The cross validation justifies the generalizability of our classifiers. The evaluation results show that our hybrid candidate extraction method lowers the chance of miss- detection and the proposed featureless classification approach has a high recognition precision. Our algorithm is robust and efficient, which can run at 30 fps for images with a resolution of 640*480.