【作者】 刘怡光；

【导师】 游志胜；

【作者基本信息】 四川大学 ， 计算机应用， 2004， 博士

【摘要】 车辆识别是智能交通系统的基础课题之一，由于车辆图片背景复杂多变、车辆外形种类繁多，因此车辆识别是一个一直没有完全解决的问题。而车辆识别具有广泛的应用，其研究受到普遍的关注，本文正是在这一背景下，对车辆识别的若干基础算法和识别技术进行了研究，主要的研究结果和创新点如下： (1)首先进行了车辆识别基础算法的研究：神经网络具有并行快速的特点，可用神经网络进行车辆识别，而要设计稳定收敛的神经网络，就要牵涉神经网络的稳定性收敛性研究。故分析了以下两部分内容：含时延和扰动的双向联想记忆神经网络的稳定性和收敛性、受扰的时滞Hopfield神经网络的稳定性。车辆的主成分提取对车辆识别是有应用价值的，结合神经网络并行快速的特点，为了用神经网络提取车辆图像的主成分，设计了一种稳定的计算矩阵特征值特征向量的神经网络。车辆识别的计算量一般较大，而遗传算法具有并行优化的特点，因此本文对遗传算法也进行了研究，提出了一种由种群发育约束个体变异的鲁棒遗传算法。 (2)构建了一种用于车辆识别的多层前向神经网络，给出了网络拓扑结构，并成功地把该神经网络运用到车型识别中。该方法综合了神经网络、模糊逻辑、模式识别的相关算法，对车辆目标轮廓进行整体识别，达到了较高的目标识别准确率。实践表明，该网络经过监督学习后，能屏除车辆中一定量干扰象素影响，准确地识别出各种外形车的车型。 (3)设计了一种可用于车辆识别的基于支持向量的多类分类器。用相异非零正实数表示不同样本模式，通过非线性映射把样本映射到特征空间，在特征空间中确定一变量系数及偏置待定的线性映射表示样本与模式实数间对应关系，映射输出为零所确定的面作为参考基准。在保证映射输出与样本模式实数四川大学博士学位论文尽量接近约束下，增大不同模式样本相对参考基准面的距离差，该目标对应优化函数与支持向量机(SvM)用于回归估计(SvR)求解的目标函数形式一致，故可类似SVR求法得到映射各待定参数。对模式实数确定方法的分析与实验表明:在学习过程确定模式实数时，该方法具有较高分类准确率。用I犯S数据对该分类器进行测试，结果显示:分类准确率优于LvQ神经网络方法，同其它SvM多类分类法比，识别速度快，准确率较高，接近DAGSvM(按照决策定向非循环图组织svM的多类分类器)准确率。 (4)建立了一种基于小波分解和Fisher变换的车辆识别方法。该方法首先对识别车辆进行多尺度小波分解，把不同尺度的分解结果组织为分解向量，再对分解向量作Fisher变换。在Fisher变换域内依据最小绝对距离(或相对距离)识别车辆。该方法主要依据小波分解低频分量进行识别运算，对噪声不敏感;同时识别结果已经融合多尺度的识别运算，故目标车辆大小对识别结果影响较弱。实践证明，该方法正确识别率高，鲁棒性强。 (5)提出了一种快速的车辆识别方法。融合小波分解与遗传算法并行性的优点，构建了一种快速鲁棒的车辆识别算法。首先对车辆和模板进行小波分解，用低频部分代替车辆和模板参与识别运算;把对模板进行形变旋转处理的参数编码为染色体，依据染色体表现值对模板进行形变和旋转处理;用模板匹配法判定车辆图片中是否含有形变旋转后的模板，判定函数就是遗传算法的适应度函数。算法终止时，最佳适应度为车辆中是否包含模板所示车辆目标的判定依据。实验表明:该车辆识别方法适应性广，稳定性好，速度快。 最后，对本文提出的各种车辆识别方法从计算复杂性、内存消耗量以及普遍适用性三方面进行了理论分析比较，同时在实验环境统一条件下，进行了实验比较分析。在综合各车辆识别方法基础上，提出了有关车辆识别技术的研究展望。关键词:车辆识别;神经网络;支持向量机;小波分解;遗传算法;—IV—更多还原

【Abstract】 Vehicle recognition is one of the basic problems in intelligent traffic system, because the background of vehicle image chops and changes and the vehicle outer shape is various, so, vehicle recognition is an unsolved problem. Its research attracts many authors, therefore, the vehicle recognition problem is studied. The main results and innovations can be described as follows:(1) First, the basic algorithms are studied: Neural network which can be used for vehicle recognition has parallel and quick prospect, in order to design stable and convergent neural network, the stable and convergence of neural networks need to be studied. So, the following two contents are received: stability and convergence research of bi-directional associative memory neural networks with perturbations and delays, The Stability of Perturbed Hopfield Neural Networks with Time Delays. Primary content analysis (PCA) has applicable value to vehicle recognition. For the parallel and quick feature, in order to extract the primary content of image, a neural network for extract the matrix eigenvalue and eigenvector is designed. Generally speaking, the computing weight for vehicle recognition is very heavy, the genetic algorithm can reduce the computing weight, therefore, a robust genetic algorithm with mutation parameters bounded by population’s upgrowth is proposed.(2) A forward multi-layer neural network model that can be used for vehicle recognition is proposed. The model’s network topology is presented, and this method is used in vehicle recognition successfully. The model integrates some results of neural network, fussy logic, pattern recognition etc, and recognize the object as a whole. The practice in the vehicle recognition using this model proofs that through supervised training, this model can correctly recognize the vehicle in spite of some interferential pixels.(3) A multiclass classifier that can be used for vehicle classification using support vector is designed. Different positive real numbers are used to denote different patterns. Through nonlinear mapping, the inputed samples are mapped into feature space, in this space, a linear mapping formula with unknown coefficient and bias is given, the nonzero output of the formula is correspondent to the pattern real number of the inputed sample. The plane corresponding to zero output of the formula is thought as reference plane. Under the condition of doing the best to minimize the difference between the linear mapping output real number and the true real pattern number of the input sample, the best are tried to maximize the distance (between the sample and the reference plane) between different patterns, the objective function of this goal is analogous to the objective function of support vector regression (SVR) formally, so, all the methods applied to SVR can be used for solving this problem, therefore, the coefficients and bias of the linear mapping formula are obtained. Through analyzing and testing the ways used to decide the pattern real number, we find that a better recognition right rate is received through deciding the pattern real number in the training process. Using the well-known IRIS data set to test the classifier,the practice indicates that the performance of this classifier is better than other methods based on LVQ neural network generally. Comparing with other methods such as 1-v-l, 1-v-r and decision tree method (DAGSVM) based on SVM, the running is quick, and the recognition right rate is high, approximately near to that of DAGSVM.(4) A new multiresolution vehicle recognition method is introduced. At first, it makes multi-wavelet division over the image, organizes the corresponding components of different resolution into a vector. Then, make a Fisher transform over the vectors. Recognition is based on the sum of the absolute distance between the vector of the unknown object and the vector of the template or based on the relative distance which is defined as the absolute distance between the vector of the unknown object and that of the template divided by the distanc更多还原