混合交通流两轮车辆的视频检测研究

发布时间：2018-03-12 08:21

本文选题：两轮车辆　切入点：前景提取　出处：《江西理工大学》2014年硕士论文　论文类型：学位论文

【摘要】：交通流的数据分析和研究是智能交通系统研究的重要组成部分,对于交通系统的安全、便捷的运行不言而喻。作为智能交通系统一部分的车辆检测也因此成为了研究的热点和重点,并取得了很多广泛应用的成果。本文从车辆检测的方向出发,结合国内外的研究现状以及国内的交通流状况,使用图像处理和机器学习的方法对两轮车辆的检测技术进行研究,采用基于模板匹配和预检测结合机器学习的方法进行两轮车辆检测,具体的研究内容是:(1)研究和总结了国内外对普通车辆以及两轮车辆检测技术,并结合实际的试验场景,对传统的车辆检测技术(如地磁线圈、超声红外传感等)和基于视频的车辆检测技术(如光流法、帧差法、背景差法)等技术的局限性、安装便捷性、数据处理的直观性进行分析和对比。(2)采用前景中两轮车辆的均值模板对两轮车辆进行检测。首先对常采用图像消噪技术如中值滤波、高斯滤波进行实验分析说明,通过对帧差法获得的前景信息进行多次形态学的膨胀处理和混合高斯模型的前景信息进行与操作,以获得更完整、噪声信息更少的前景。使用边缘检测的方法获得运动图像中的车辆边缘信息,并利用前景中两轮车辆的均值获得模板,并反馈到前景中,与前景中的动态目标进行模板匹配,检测两轮车辆。利用车辆质心轨迹分析的方法对检测的车辆进行计数。(3)运用混合高斯模型和Ada Boost算法进行车辆检测。检测步骤包括:利用小汽车和两轮车辆的形状特征的不同性进行预检测,使用预先获取的正样本和负样本以及Ada Boost机器学习的方法对样本的LBP、HAAR、HOG特征分别进行分类器的训练。并使用分类器在预检测的窗口上进行两轮车辆的检测,通过训练时间以及训练得到的分类器在视频序列中的检测的正确率的分析,得出最符合本文环境的检测特征,即LBP特征。实验表明,本文提出的在预检测的基础上使用机器学习进行两轮车辆检测的方法可以明显加快检测速度,并有效降低误检率。
[Abstract]:Traffic flow data analysis and research is an important part of intelligent transportation system research. The convenient operation is self-evident. As a part of the intelligent transportation system, vehicle detection has become the focus and focus of the research, and has made a lot of widely used results. This paper starts from the direction of vehicle detection. Combined with the domestic and foreign research situation and the domestic traffic flow situation, using the image processing and the machine learning method to carry on the research to the two-wheeled vehicle detection technology, The two-wheel vehicle detection method based on template matching and pre-detection combined with machine learning is adopted. The specific research content is: 1) the research and summary of the domestic and foreign common vehicles and two-wheel vehicle detection technology, and combined with the actual test scene, It is convenient to install traditional vehicle detection technology (such as geomagnetic coil, ultrasonic infrared sensor, etc.) and video-based vehicle detection technology (such as optical flow method, frame difference method, background difference method). The visual analysis and contrast of data processing. (2) using the mean value template of the two wheel vehicle in the foreground to detect the two wheel vehicle. Firstly, the image denoising technology such as median filter and Gao Si filter are used for experimental analysis. The foreground information obtained by frame difference method is processed by morphological expansion several times and the foreground information of mixed Gao Si model is processed and operated in order to obtain a more complete picture. The edge detection method is used to obtain the vehicle edge information in the moving image, and the template is obtained by using the mean value of the two-wheeled vehicle in the foreground, which is fed back to the foreground and matched with the dynamic target in the foreground. Two-wheeled vehicles are detected. The vehicle is counted by the method of centroid trajectory analysis. The hybrid Gao Si model and Ada Boost algorithm are used to detect the vehicle. The detection steps include: using the shape of the car and two-wheeled vehicle. The different characteristics of the character are pre-detected. Using pre-acquired positive and negative samples and Ada Boost machine learning method, the classifier is trained for the LBPHAARHOG feature of the sample, and the classifier is used to detect the two-wheeled vehicle on the pre-detected window. Through the analysis of the training time and the correct detection rate of the classifier in video sequence, the LBP feature, which is the most suitable for the environment of this paper, is obtained. The method of two-wheel vehicle detection based on machine learning proposed in this paper can significantly accelerate the detection speed and effectively reduce the false detection rate.
【学位授予单位】：江西理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495;TP391.41

【参考文献】