基于支持向量机的路面状态视频图像识别技术研究

发布时间：2018-03-30 05:02

本文选题：高速公路　切入点：视频图像　出处：《北京交通大学》2017年硕士论文

【摘要】：恶劣的路面条件是造成道路交通事故的重要诱因,当车辆在雨、雪、雾等恶劣天气下行驶时,极易发生车辆打滑侧翻等重特大交通事故。因此,准确快速地对路面状态进行检测识别,并在道路条件恶劣的情况下做出及时的预警和反应,对高速公路的高效安全运营具有重大现实意义。随着视频图像处理技术的成熟发展及高速公路全程视频监控系统的普及,监控摄像机已成为主要的交通监测设施。因此,利用视频图像技术识别路面状态已成为当前研究热点。然而,对于混合路面状态的识别,以及不同光照条件下的路面状态识别是两个亟需重点解决的难题。本论文提出利用视频图像技术检测路面状态,并开发可满足不同路段需求的路面状态检测算法。具体研究内容如下:(1)论文首先将路面状态分为干燥、潮湿、积雪、结冰、积水5种状态,根据原始图像大小,制定了图像分块原则,提取单一状态的图像块构建了路面状态图像库,保证了样本的质量和纯度。(2)采用三阶颜色矩法提取了 9维颜色特征向量,采用灰度共生矩阵法提取能量、熵、相关性、对比度4个纹理特征量,基于该13维图像特征向量,形成了路面状态特征数据库。(3)提出了基于SVM(支持向量机)的路面状态视频识别方法,为提升该算法的普适性的识别精度,采用网格搜索算法优化了 SVM中的核函数因子C和惩罚因子g。(4)建立路面状态图像分类模型,首先,利用参数寻优的SVM分类器对多组不同样本量进行了训练,得到多组路面状态图像分类模型;接着,对上述多组分类模型进行性能测试识别,甄选出分类效果最佳的模型。(5)依托搭建的实验系统及源于不同采集方式的视频样本,对大量不同环境、未经训练的实际路面状态原始图像进行了分块识别验证。实验结果表明:基于SVM寻优分类器和视频图像分块识别的方法科学可行,网格搜索寻优算法下的路面状态分类模型对单一状态下的样本识别准确率90%以上,对混和路面样本的识别准确率85%以上。有效解决了混合状态路面状态和不同光照条件下路面状态的识别难题。
[Abstract]:Bad road conditions is an important cause of road traffic accidents, when the vehicle is in the rain, snow, fog and other inclement weather driving, prone to skid rollover and other serious traffic accidents. Therefore, accurate detection of road conditions, and make early warning and timely response in a bad way the case is of great practical significance to the safe operation of the highway, and highway. With the popularity of mature development of the whole video monitoring system of video image processing technology, surveillance cameras have become the main traffic monitoring facilities. Therefore, the use of video image recognition technology of pavement state has become a hotspot of current research. However, for the identification of mixed pavement state, and the road condition identification under different illumination conditions are two need to focus on resolving the problem. This paper presents the use of video image detection technology Road conditions, and development can meet the needs of different sections of pavement state detection algorithm. The specific contents are as follows: (1) the first road state is divided into dry, wet, snow, ice, water in 5 states, according to the size of the original image, formulated the principle of image block, image block extraction single state construction road condition image library, to ensure the quality and purity of the sample. (2) the extraction of 9 dimensional color feature vector by three order color moment method using gray level co-occurrence matrix method to extract energy, entropy, correlation, contrast 4 texture features, the 13 dimensional image feature vector based on the formation of the State Road feature database. (3) based on SVM (support vector machine) the road condition identification method for video, boosting the algorithm's general recognition accuracy, using the grid search algorithm to optimize the function of nuclear factor C g. and the penalty factor in SVM (4) established State road image classification model, firstly, using SVM classifier parameter optimization was trained on different samples, resulting in multiple sets of pavement state image classification model; then, test the performance of recognition to the group classification model, the selection of the best results of classification model. (5) based on the experimental setup and due to the different acquisition methods of video sample, a large number of different environment, without the state of actual pavement training image block identification. The experimental results show that the optimization method based on SVM classifier and video image block recognition is scientific and feasible, grid search algorithm under the pavement state classification model for sample identification under the single state accuracy rate above 90%, the recognition accuracy of the sample mixture pavement above 85%. Effective solution to the identification of mixed state of pavement state and under different illumination conditions of pavement state Difficult problems.

【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U418.6

【参考文献】