在智能交通系统(简称ITS)中，车辆检测和车型识别问题作为其中的关键技术一直备受研究者们的关注。该技术旨在准确的检测和统计车辆数量，同时为现有的车辆检测系统增加类型信息，从而改善交通阻塞状况。与此同时，伴随着传感器技术和人工智能技术的发展，越来越多的研究者致力于研究模式识别算法解决车辆检测和车型识别问题，并将取得的研究成果应用到ITS中。在这些研究工作的基础上，本篇论文基于各向异性磁阻式传感器(AMR)采集来的磁扰动信号，针对车辆检测和车型识别问题展开了进一步的研究，重点讨论了车辆检测干扰问题、车型识别特征提取问题和车型识别Class-Imbalance问题，为ITS中车辆检测和车型识别等问题研究提供了有效的理论支撑。 首先，针对车辆检测容易受到相邻车道上车辆干扰的问题，本文提出了一种基于BP神经网络的车辆检测方法，有效消除了上述干扰因素的影响。通过对实际交通场景的细分，本文在研究工作中将干扰源划分为远车道状态、重叠状态以及噪声状态，并对目标状态和干扰项构建出了有效的特征集合。仿真结果显示，通过使用提出的BP神经网络车辆检测方法，能获得高达99.51％的车辆检测准确率，解决了车辆检测过程中的干扰问题。 其次，针对传统车型识别方法准确率不高的问题，本篇论文提出了一种时频域特征融合的方法，有效的提升了车型分类器的性能。在时域特征方法的设计过程中，本文分别从磁信号的极值、最值、全局特征和能量的角度构建了如下五种特征集合：Hill-Pattern特征集，Peak-Peak特征集，Mean-Std特征集，Energy特征集以及Time特征集(前四种特征集合的融合)。接下来，通过对磁信号频谱的研究，提出了在频域有效的特征提取方法。最后，通过融合设计的时域方法和频域方法，获得了一种有效的车型识别算法。仿真结果显示，基于时频域特征方法构建的特征集合，使用BPNN算法能获得最高81.82％的分类准确率。值得注意的是，本文讨论的车型分类问题主要针对如下四种车型：两厢车、三厢车、大巴车以及多用途汽车。 最后，针对车型识别的Class-Imbalance问题，本文在传统的过取样方法基础上，提出了一种基于SMOTE方法改进的KNN算法，该算法从数据集重构的角度有效的解决了上述问题。在实际的交通情况下，Class-Imbalance问题是指由不同车型数量差异较大带来的误差问题，该问题会导致预测的类型严重偏向于交通流中占比大的车型。本文通过研究不同车型引起的磁扰动信号特征，发现了由同种车型引起的特征呈现了聚类的特性，从而得出SMOTE算法适用于车型分类识别应用的结论。仿真结果显示，基于本文提出的时频域特征集合，使用基于SMOTE方法改进的KNN算法不仅有效的解决了Class-Imbalance问题，而且能获得高达95.46％的车型分类准确率。 关键词：智能交通系统，车辆检测，车型分类识别，地磁传感器

 In Intelligent Transportation System (or ITS for short), vehicle detection and vehicle classification as one of the key technologies has been the concern of the researchers. The technology aims to accurately detect and statistics the number of vehicles, and increase the type of vehicle information for an existing vehicle detection system, thereby improving the existing traffic conditions. At the same time, with the development of sensor technology and artificial intelligence technology, more and more researchers are working on pattern recognition algorithms to solve vehicle detection and vehicle classification problems, and apply the research results to ITS. Based on their work, this paper has carried out further research on removing interference in vehicle detection, the design of feature extraction methods and the "Class-Imbalance" problem by using anisotropic magnetoresistive sensor (AMR), which provide effective theoretical support for the research of vehicle detection and vehicle classification in ITS. Firstly, in order to solve the problem that vehicle detection is easily affected by the vehicle on the adjacent lane, this paper proposes a vehicle detection method based on BP neural network, which effectively eliminates the influence of the above-mentioned interference factor. Through the subdivision of the actual traffic scene, this paper divides the interference source into "Far" state, "Overlapping" state and "Noise" state, and constructs an effective feature set for the target state and the interference term. The simulation results show that by using the proposed BP neural network vehicle detection method, the vehicle detection accuracy of up to 99.51% can be obtained, and the interference problem in the vehicle detection process is solved. Secondly, due to the accuracy of the traditional vehicle classification method is not high, this paper proposes a method of time domain and frequency domain features fusion, which effectively improves the performance of the vehicle classifier. In the design process of the time domain feature method, this paper constructs the following five feature sets in the time domain, which are Hill-Pattern feature set, Peak-Peak feature set, Mean-Std feature set, Energy feature set and Time feature set (consists of the above four feature sets). Next, through the study of the magnetic signal spectrum, a feature extraction method effective in the frequency domain is proposed. Finally, an efficient vehicle classification algorithm is obtained by integrating the time domain method and the frequency domain method. The simulation results show that the feature set constructed based on the proposed method can obtain the classification accuracy of up to 81.82% using BPNN algorithm. It is worth noting that the classification of the models discussed in this article is mainly for the following four models: hatchback, sedan, bus and multi-purpose vehicles. Finally, in view of the problem of Class-Imbalance in vehicle classification, this paper proposes an improved KNN algorithm based on the traditional oversampling method. This algorithm effectively solves the above problems from the perspective of data set reconstruction. In the actual traffic situation, the Class-Imbalance problem refers to the error caused by the large difference in the number of different models. This problem will lead to a serious bias in the type of forecasting in the traffic flow. By studying the characteristics of magnetic disturbance signals caused by different models, we find that the characteristics caused by the same type of vehicle presents clustering characteristics, and conclude that the SMOTE algorithm is suitable for vehicle classification applications. The simulation results show that based on the feature set obtained in this paper, the SMOTE-KNN algorithm not only effectively solves the Class-Imbalance problem, but also can obtain the classification accuracy of up to 95.46%. Key Words: Intelligent Transportation System, Vehicle Detection, Vehicle Classification, Magnetic Sensor