Traffic Signs Recognition Based On Sparse Representation

Traffic sign recognition as an important part of the Intelligent Transportation Systems, has become a hot issue in the field of vision research at home and abroad. A wide range of traffic signs, faced with complex background of natural scenes, light pollution, deformation, occlusion and other factors, these traffic sign recognition method accuracy, robustness and real-time put forward higher requirements. In recent years, sparse representation based classification method has a high recognition rate, robustness advantages been more and more attention, and get better experimental results in face recognition. This paper focuses on the traffic sign recognition method based on sparse representation, Research and improvement of the critical steps.This paper first make the image-preprocessing of the experimental samples of German traffic sign database. Then analysis and comparison the experimental result of classical sparse decomposition algorithm-orthogonal matching pursuit algorithm. Orthogonal matching pursuit algorithm also has high recognition accuracy and robustness in traffic sign recognition.Then we describe and analyze a new sparse representation method a two-phases sparse recognition algorithm. The algorithm in small-scale face recognition has a better experimental result. But if the number of categories and the training sample set is large, the algorithm strike sparse representation need for a very large memory overhead, and therefore there is a certain limit in actual use. According to the insufficient of the algorithm, we split redundant dictionaries into multiple local dictionaries, and select the M nearest neighbors from the local dictionaries. Experimental results show that the proposed method in traffic sign recognition has a better recognition effect than the traditional orthogonal matching pursuit algorithm.The M nearest neighbors which elected by local dictionaries are not overall.So we proposed recognition algorithm based on the sparse representation of kernel distance. The algorithm uses distance function which is induced by Gaussian kernel function to measure the similarity of the test sample and the training sample, from which to sparse represent the test samples using M nearest training samples. Kernel distance reflected in the training sample differences very well, so the recognition accuracy of this algorithm has been further improved. The experimental results show that sparse recognition algorithm based on kernel distance has higher recognition rate, and enhance the computing speed.A wide variety of traffic signs, many of which identifies the presence of similar, these similar identification can easily be mistaken for classification. In response to this situation, we propose a traffic sign recognition method based on two-layers of crude and fine as the improvement of the sparse representation algorithm based on kernel distance. First we divided the similar traffic signs category into three main categories, which is the rate-limiting class, warning class and the others; Then make a crude classification based on the kernel distance sparse algorithm; Finally, we make a fine classification on the rate-limiting class and warning class. The experimental results show that the method based on two-layers of crude and fine is better than the method directly gives the recognition result, recognition accuracy rate of about1%.