Research Of Vehicles Recognition Methods Based On Sparse Representation

Although after long time of development, research of vehicle classification and recognition is still a very important research topic in scientific research institutions at home and abroad. In particular, vehicle detection based on image analysis is becoming a hot spot of academic and industry to focus on due to the low cost, high flexibility and can avoid potential accidents. But because there are still technical problems unable to be solved at this stage, such as the moving vehicles in natural scenarios due to the existence of background change and part of the shelter, it is not easy to identify, etc. Therefore, it is necessary to expand related theory research.Vehicles under the natural scenarios always have inconsistency because of the change of the weather and light conditions, and the vehicle itself has difference in such aspects as color, size, and posture; in addition, under the natural scene, the shelter problems, some even serious barriers, are inevitably kept out, this problem no doubt increases the difficulty of the vehicle identification. This dissertation researches in view of vehicle identification and classification problem under the natural scene and does the following several aspects:(1) First, this dissertation introduces the definition of the image classification and target recognition summarily, and then summarizes the research background and significance of the vehicle identification and classification, reviews the development status of vehicle identification at home and abroad, makes the focus of the study in this paper clear;(2) Next, the dissertation presents the origin and the development of the sparse representation and several popular sparse representation algorithms, such as global optimization algorithm, greedy algorithm, and other categories, and analysis the advantages and disadvantages of various algorithms at the same time;(3) In view of vehicle identification problem of the natural scene, this dissertation uses a suitable feature extraction method, first uses PCA dimension-reduction method to simplify the dimensions of the image, and then SIFT is used to extract the invariant scale. At the same time, the feature matching problem is introduced and solved;(4) Then, this dissertation puts forward a kind of K-SVD dictionary training based on kernel function combined with sparse representation classification method:first of all, after the use of the described feature extraction method on image processing, we use the kernel function on the characteristic matrix and the matrix is mapped into a high dimensional space from a lower dimensional space, in this new high dimensional space we train new high-dimensional feature matrix with K-SVD into two kinds of characteristics of the dictionaries. At this point, the sparse representation classification SRC can be used on test sample’s classification and identification. It is found that after compared with several other classical identification methods, the presented method can improve the recognition rate, more importantly, is capable to eliminate partial sheltering effect on vehicle identification;(5) Finally, this dissertation proposes to use today’s popular classifier—support vector machine (SVM) combined with sparse representation methods to ensure the accuracy of recognition rate. Firstly introduces the definition and classification theory of support vector machine, analyzes the advantages and disadvantages; Then an algorithm is designed to combine the two methods and proposes relevant experimental verification, the experimental results show:the combination on the one hand, not only inherits the advantages of SVM—to get the global optimal solution in the identification process, and avoid its disadvantages:Reduce the expense of SVM on the operation cost, at the same time, also ensures that the recognition performance obtained by the sparse representation classifier and keep the advantage of the sparse representation without prior training, reducing the amount of calculation greatly.The end of the dissertation summarizes the research content and analyzes the experimental results and performance in this paper, puts forward the idea of rationalization aspects which can improve and perfect the results, and finally makes the further discussion and outlook of vehicle identification and classification.