Speed Limit Sign Recognition Using Log-Polar Mapping and Visual Codebook

Abstract

Traffic sign recognition is one of the hot issues on the modern driving assistance. In recent years, the method using Bag-of-Word (BOW) model for image recognition has gained its popularity upon its simplicity and efficiency. The conventional approach based on BOW requires nonlinear classifiers to get a good image recognition accuracy. Instead, a method called Locality-constrained Linear Coding(LLC) presents an effective strategy for coding, and only with a simple linear classifier could achieve a good effect. LLC uses uniform sampling for feature extraction, but allowing for features of traffic signs, the central vision information of the image is more important than the surroundings. Fortunately, log-polar mapping to preprocess image samples before coding is helpful for traffic sign recognition. In this paper, a combination method of log-polar mapping and LLC algorithm is presented to achieve a high image classification performance up to 97.3141% on speed limit sign in the GTSRB dataset.