基于稀疏表示的车辆识别方法研究

发布时间：2018-04-21 05:45

本文选题：车辆识别 + 主成分分析　；参考：《合肥工业大学》2014年硕士论文

【摘要】：车辆的分类与识别的研究历经长时间的发展,至今仍是国内外众多科研机构的重要研究方向。特别是基于图像分析的车辆检测,由于其成本低、灵活度高以及可以避免潜在的交通事故,成为学术界和工业界争相关注的热点。但是由于依然存在很多现阶段无法解决的技术难题,如自然场景下的运动车辆由于存在背景变化和部分遮挡而不易识别等,因此,有必要展开相关的理论研究工作。自然场景下的车辆,往往都存在着天气、光照条件的变化以及车辆本身在颜色、尺寸、姿态等方面的不一致；另外,自然场景下不可避免的是遮挡问题,有些甚至是严重遮挡,这无疑加大了车辆识别的难度。本文针对自然场景下的车辆识别与分类问题,做了以下几方面的工作：(1)首先概括的介绍了图像分类与目标识别的基本理论,然后总结了车辆识别与分类的研究背景和意义,回顾了国内外车辆识别的发展现状,明确了本文所要研究的内容；(2)介绍了稀疏表示的发展由来以及稀疏表示的几种流行算法,如全局优化算法、贪婪算法以及其他几类算法,同时分析了各种算法的利弊；(3)针对自然场景下的车辆识别问题,利用一种适合的特征提取方法,先用PCA降维方法对图像的维数进行简化,接着用SIFT提取出尺度不变性特征。同时还描述了特征匹配问题；(4)提出一种基于核的K-SVD字典训练结合稀疏表示的分类方法：首先在利用上述所介绍的特征提取方法对图像进行预处理之后,然后对所得特征矩阵利用核函数将其从低维空间映射到高维空间,在这个高维空间里用K-SVD对新的高维特征矩阵进行训练,分别得到相应的两类特征字典。此时,即可用稀疏表示分类器SRC对测试样本进行识别分类。在与其他几种经典识别方法对比后发现,所提出的方法识别率有改善,更重要的是能够消除部分遮挡对车辆识别的影响；(5)最后,提出利用当今比较流行的分类器——支持向量机SVM与稀疏表示结合的方法来确保识别率的准确性。首先介绍了SVM的理论及分类原理,分析了其优缺点；接着用二者结合的思想进行了算法设计和实验验证实验结果证明：这种结合一方面不仅继承了SVM的优点——识别过程中得到的是全局最优解,并且避免了其缺点：减少了SVM在运算消耗上的代价同时结果逼近由稀疏表示分类器所得的识别性能,且保留了稀疏表示无需事先训练的优势,大大减少了计算量。最后总结了本文所研究内容,分析了实验的结果和性能,对可以提高和完善的方面提出了合理的构想,并对车辆识别与分类的发展做出了进一步的探讨和展望。
[Abstract]:The research of vehicle classification and recognition has been developing for a long time, and it is still an important research direction of many scientific research institutions at home and abroad. Especially, vehicle detection based on image analysis, because of its low cost, high flexibility and can avoid potential traffic accidents, has become a hot spot in academia and industry. However, there are still many technical problems that can not be solved at this stage, such as the moving vehicles in natural scenes are difficult to recognize because of background changes and partial occlusion, so it is necessary to carry out relevant theoretical research work. In the natural scene, the vehicle often has the weather, the illumination condition change and the vehicle itself in the color, the size, the posture and so on inconsistent; in addition, under the natural scene inevitably is the occlusion question, some is even the serious occlusion, This undoubtedly increases the difficulty of vehicle identification. Aiming at the problem of vehicle recognition and classification in natural scene, this paper firstly introduces the basic theory of image classification and target recognition, then summarizes the research background and significance of vehicle recognition and classification. This paper reviews the development of vehicle recognition at home and abroad, and clarifies the content of this paper.) the development of sparse representation and several popular algorithms of sparse representation, such as global optimization algorithm, greedy algorithm and other algorithms, are introduced. At the same time, the advantages and disadvantages of various algorithms are analyzed. Aiming at the problem of vehicle recognition in natural scene, a suitable feature extraction method is used to simplify the dimension of the image by using PCA dimensionality reduction method, and then to extract the scale invariant feature by using SIFT. At the same time, the feature matching problem is also described. A kernel based K-SVD dictionary training method combined with sparse representation is proposed. Firstly, after preprocessing the image by using the feature extraction method mentioned above, Then the feature matrix is mapped from low-dimensional space to high-dimensional space by kernel function. In this high-dimensional space, the new high-dimensional feature matrix is trained by K-SVD, and two kinds of feature dictionaries are obtained respectively. At this point, the sparse representation classifier SRC can be used to identify and classify the test samples. Comparing with other classical recognition methods, it is found that the recognition rate of the proposed method is improved and, more importantly, the effect of partial occlusion on vehicle recognition can be eliminated. In order to ensure the accuracy of recognition rate, a new method combining support vector machine (SVM) and sparse representation is proposed. Firstly, the theory and classification principle of SVM are introduced, and its advantages and disadvantages are analyzed. Then, the algorithm is designed and the experimental results show that the combination not only inherits the advantages of SVM, but also obtains the global optimal solution in the process of recognition. And the disadvantages are avoided: the cost of SVM is reduced and the result approximates the recognition performance obtained by the sparse representation classifier, and the advantages of sparse representation without prior training are preserved, and the computational complexity is greatly reduced. Finally, the contents of this paper are summarized, the results and performance of the experiment are analyzed, a reasonable conception is put forward for improving and perfecting, and the development of vehicle recognition and classification is discussed and prospected.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41;U495

【相似文献】