Study On Large Category Traffic Sign Recognition Algorithm

Automatic traffic sign recognition (TSR) not only plays an important role in intelligent transportation systems (ITS), but also is an essential component of unmanned vehicles. When applied in advanced driver assistance systems (ADAS), traffic sign recognition can provide the drivers information about road traffic conditions, traffic regulations, driving directions etc., so as to improve traffic safety and reduce traffic accidents. Traffic sign recognition in complex natural scene is vulnerable to weather conditions, illumination changes, motion blur, partial occlusion, rotation, etc. so the rapid and accurate traffic sign recognition meets a considerable challenge.Traffic signs can be divided into a number of super categories which contain many sub-categories. Traffic signs appearances in different super categories are very different from each other but the sub-categories in a same super category are very similar to each other. In order to improve the accuracy and speed of traffic sign recognition, this thesis proposed a coarse-to-fine hierarchical classification strategy:firstly, we use a single feature (improved histogram of oriented gradient feature) and support vector machine (SVM) to classify a traffic sign into its super category, then local binary pattern (LBP), Gabor wavelet and dense SIFT are extracted, fused, and fed to a committee of support vector machine and random forest to recognize the traffic sign’s final class.The main works in this thesis are as follows:1. Original histogram of oriented gradient based on gray image can’t effectively distinguish traffic signs with similar texture but different colors. In this thesis, we extract HOG features from multiple channels of the color traffic sign images. The improved HOG features contain color information of traffic signs. As a result, the classification performance is improved.2. As a single feature is difficult to fully describe the characteristics of the traffic signs, this thesis deeply studied three local feature descriptors:local binary pattern, Gabor wavelet and dense SIFT. These three features are fused to enhance the ability to distinguish similar traffic signs.3. This thesis deeply studied support vector machine and random forest (RF). In order to be more effective to distinguish traffic signs with similar appearance, support vector machine and random forest are fused, and as a result the classification accuracy is improved significantly. Finally, we test the performance of the proposed method on German traffic sign recognition benchmark (GTSRB). Experimental results show that the proposed method meets the accuracy, robustness and real-time requirements for traffic sign recognition system.