【作者】 朱方；

【导师】 杨瑞霞； 顾军华；

【作者基本信息】 河北工业大学 ， 微电子学与体电子学， 2010， 博士

【摘要】 客流数据是整个公交企业管理业务的基础,快速、准确地采集车辆的客流信息为科学合理地安排调度车辆、优化公交线路等智能管理提供了最基本的依据,还可以全面如实地反映公交车辆的实际载客人数,方便与钱箱收入之间的核对。本文介绍了目前比较普遍的客流识别方法,并总结了优缺点,提出将多信息融合技术运用到客流识别中来。在信息科学技术领域中,多源信息融合是一个有广泛应用背景及重要理论意义的研究课题。常用的信息融合算法有加权法、Bayes法、证据组合理论、模糊逻辑、神经网络等。这些方法大多依赖于先验知识,从而造成在小样本、高维空间情况下出现模式识别效果不佳的问题。为了解决这个问题本文将支持向量机引入到多信息融合模式分类中来,并对于支持向量机算法进行了研究。根据客流识别的实际问题对于支持向量机的训练算法与快速分类算法进行了改进。最终将多信息融合技术引入到客流识别领域中,构建了基于支持向量机的多信息融合模型,应用嵌入式技术设计并实现了多功能信息采集车载终端。首先针对于支持向量机在实际训练中由于学习样本集很大且具有类内混杂孤立点数据,引发的学习速度过慢、存储需求量大、泛化能力降低等问题,本文根据点集理论提出了一种新的支持向量机大规模训练样本集缩减策略(SVM-LSTSRS)。该策略运用模糊聚类方法去除类内孤立点并提取出潜在支持向量,从而有效的避免孤立点数据所造成的过学习现象,提高SVM学习机器的泛化性能,并且大大减少训练样本集的规模,在不降低分类精度的前提下提高训练速度。最终通过实验验证了该方法的有效性和可行性。其次针对支持向量机(SVM)分类速度取决于支持向量数目的应用瓶颈,提出一种SVM快速分类算法。通过引入支持向量在特征空间的相似性度量,构建特征空间中的最小支撑树,在此基础上将支持向量按相似性最大进行分组,依次在每组中找到决定因子和调整因子,用两者的线性组合拟合一组支持向量在特征空间的加权和,从而减少支持向量的数量,提高支持向量机的分类速度。实验结果证明,该方法能以很小的分类精度损失换取较大的分类时间缩减,满足支持向量机实时分类的要求。最终通过大量实验对人体上下台阶时的压力变化规律进行了研究。提出在单人情况下,可以应用压力数据的特征进行模式分类的方法。针对于双人情况,加入时序信息,提出应用多信息融合的模式分类方法,建立了应用于客流识别的多信息融合模型,并构造了模式分类系统。通过实验表验证了系统的可行性。根据公交行业的现实需求,本文设计并实现了嵌入式客流信息采集车载终端。更多还原

【Abstract】 The passenger flow data is the basis of public transportation management. Correct and real-time data of passenger flow can provide the most important information for reasonably dispatching buses and optimizing lines of public transportation, and it can also show the actual number of passengers for checking back the cash in the cashbox conveniently. This paper presents some general methods of bus passenger flow recognition, and generalizes the advantages and disadvantages. The multi-information fusion technology is introduced into passenger flow recognition in this paper. Multiple source information fusion has widespread application and important theory significance in the area of information science and technology. Common multi-information fusion methods include weighting method, Bayes method, evidence combination theory, fuzzy logic and neural network, etc. The recognition results are dissatisfied under the conditions of small sample and high dimension space, because most of these methods are based on prior knowledge. Support Vector Machine (SVM) is introduced into multi-information fusion pattern classification to solve this problem. According to the practical problems of passenger flow recognition, the methods to improve training efficiency and speed up classification are researched on, and some new methods are proposed in this paper. On these bases, the multi-information fusion model of passenger flow recognition is established and the passenger flow collection vehicle terminal is developed and implemented.Recent years, support vector machine has become increasingly popular tools for machine learning tasks which involving pattern recognition, regression analysis and feature extraction. Because of large-scale training samples and outlier data immixed in the other class, there are some disadvantages of support vector machine such as slow learning speed, large buffer memory requirement, and low generalization performance. Aiming at these problems, a new reduction strategy for large-scale training samples according to the point set theory is proposed in this paper. This new strategy reduces the outlier data immixed in the other class and get the support vector by using fuzzy clustering. It can effectively avoid over-learning which is caused by outlier data, improve the generalization performance of the SVM learning machine and greatly reduce the scale of training samples. So the learning rate can be speed up and the classification accuracy can be unaffected. Effectiveness and feasibility of this strategy are proved by experiments.There is a bottleneck of Support Vector Machine: the speed of classification depends on the number of support vectors. Aiming at this, a fast classification algorithm of support vector machine is proposed. Firstly, it constructs the minimum spanning by introducing the similarity measure and divides the support vectors into groups according to the maximum similarity in feature space. After then, the determinant factor and the adjusting factor will be found in each group by some rules. In order to simplify the support vectors, the linear combination of“determinant factor”and“adjusting factor”will be used to fit the weighted sums of support vectors in feature space. Finally, the speed of classification will be improved significantly. Experimental results show that this algorithm can obtain higher reduction rate of classification time with minor loss of classification accuracy. It can also satisfy the requirements of real-time classification.In this paper, large numbers of experiments are carried out to research the pressure data of getting on/off the bus. It is proposed that pressure characteristic vector can be used for pattern classification when only one person getting on/off the bus. A new passenger flow recognition method based on multi-information fusion combining pressure data with timing information is presents for two persons getting on/off the bus at the same time, and the multi-information fusion model of passenger flow recognition is established. The experimental results show that the proposed model is effective and its precision is preferable. Finally, according to requirements of practical problem, passenger flow collection vehicle terminal is developed and implemented.更多还原