Traffic Sign Recognition Based On Local Features

With the rapid development of China’s economy, more and more private vehicles come into a huge number of families. It makes people’s life more convenient, but also causes a wide range of traffic problems. It is effective to solve these problems by establishing the Intelligent Transportation System (ITS). As a part of ITS, Traffic Signs Recognition (TSR) technology plays an important role in vehicle automated driving, assistant decision-making and driving safety improving. Local feature gradually gets more and more attention, because it is with the unique advantages in scale invariance, rotation invariance and affine invariance. According to the characteristics of traffic sign, this dissertation uses local feature to describe the image and realize traffic sign recognition by image feature point matching and image classification.In the first method, SIFT (Scale Invariant Feature Transform), SURF (Speeded Up Robust Feature) and ASIFT (Affine-SIFT) are separately used to extract local features, then matching points are obtained. The category of the potential sign is determined by the number of matching points.In the second method, this dissertation proposes a method combined bag of words model with SVM (Support Vector Machine) classifier. Firstly, the local feature is extracted and described by SIFT algorithm. Secondly, the bag of words model is applied to generate PHOW (Pyramid Histogram of Words) descriptor for representing the sign. Thirdly, the LIB SVM classifier is adopted to realize the classification.The dataset which contains280signs is built by online searching and practical shooting, including standard signs, noise-added signs, rotated signs, motion blur signs, viewpoint-changed signs and partial occlusion signs. All the images are normalized to200X200.20categories signs are tested on the PC with processor2.2GHz, the experiments show that ASIFT-based method reaches over87percent success rate. The accuracy of SIFT-based method and SURF-based method are separately lower by10percent and12percent. SURF-based method cost1.37seconds on average to output the result. SIFT-based method is almost twice of it and ASIFT-based method is much longer, with about3.5times of SURF-based method. The classification method reaches91percent accuracy rate and costs about2.02seconds in processing each sign. Compared with matching method, it performs better in recognition and faster in implementation.