复杂场景下车辆（动目标）的识别和跟踪技术研究

发布时间：2018-05-31 23:11

本文选题：混合高斯模型 + 阴影消除　；参考：《南京航空航天大学》2014年硕士论文

【摘要】：在目前的智能交通系统中，对车辆的识别和跟踪一直是一个核心的环节，它能够提供各种动态的交通环境信息，便于统一管理和调度，缓解交通拥挤和减少交通事故，因此对车辆的准确识别和长期跟踪一直是智能交通监控的研究热点。本文重点研究了车辆的识别和跟踪理论，从四个步骤重点论述了车辆的检测、识别和跟踪方法，并用具体的实验证明了本文算法的可靠性和有效性。具体的工作如下：（1）提出了一种基于改进的混合高斯模型的动目标检测算法，，该算法通过帧间匹配度信息反馈改变了传统方法的学习规则，克服了车辆检测断裂或分离的缺陷，排除了车辆和环境对背景学习的干扰。实验表明，该方法对于提取动目标区域较经典方法更加准确。（2）提出了一种基于HSV色彩空间法和混合高斯模型的阴影检测算法，该算法通过人工采集方法和HSV色彩空间法来获得阴影样本，并利用期望最大法对阴影训练样本估计模型参数，获得的混合高斯模型用来区分车辆和阴影。实验结果表明该方法可以有效分离车辆和阴影。（3）采用了7个Hu不变距、分散度、长宽比和紧凑度组成10维的形状特征向量以及三层BP神经网络对行人、大车、小车、自行车或者摩托车这四类目标进行分类，实验结果表明通过样本训练出来的神经网络分类器可以对这四类目标有效分类。（4）提出了一种改进的TLD跟踪算法，该算法结合原来的单分类器，加入了基于Haar特征和在线Adaboost方法的分类器，构成了一种半监督协同训练的分类器，提高了分类器的泛化能力，实验结果表明该方法可以进一步提高跟踪效果。
[Abstract]:In the current intelligent transportation system, the identification and tracking of vehicles is always a core link. It can provide a variety of dynamic traffic environment information, facilitate unified management and scheduling, alleviate traffic congestion and reduce traffic accidents. Therefore, the accurate identification and long-term tracking of vehicles has always been the research hotspot of intelligent traffic monitoring. This paper focuses on the theory of vehicle recognition and tracking, discusses the detection, recognition and tracking methods of vehicles from four steps, and proves the reliability and effectiveness of this algorithm by experiments. The specific work is as follows: (1) A moving target detection algorithm based on the improved hybrid Gao Si model is proposed. The algorithm changes the learning rules of the traditional methods through the information feedback of the matching degree between frames, and overcomes the defect of vehicle detection breaking or separation. The interference of vehicle and environment to background learning is eliminated. Experiments show that the proposed method is more accurate than the classical method in extracting moving target regions. (2) A shadow detection algorithm based on HSV color space method and hybrid Gao Si model is proposed. The shadow sample is obtained by artificial acquisition and HSV color space method, and the model parameters are estimated by the expected maximum method. The hybrid Gao Si model is obtained to distinguish vehicles from shadows. The experimental results show that this method can effectively separate vehicle from shadow. Using seven Hu invariants, dispersion, aspect ratio and compactness, a 10-dimensional shape feature vector and a three-layer BP neural network are used to classify pedestrian, cart, car, bicycle or motorcycle. The experimental results show that the neural network classifier trained by the samples can effectively classify the four kinds of targets. In this paper, an improved TLD tracking algorithm is proposed, which combines the original single classifier and adds a classifier based on Haar features and online Adaboost method. It constitutes a semi-supervised cooperative training classifier and improves the generalization ability of the classifier. Experimental results show that this method can further improve the tracking effect.
【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495

【参考文献】