智能交通系统中行人快速检测算法研究

发布时间：2018-02-20 02:08

本文关键词： 行人快速检测智能交通系统梯度直方图 PCA LBP　出处：《成都理工大学》2016年硕士论文　论文类型：学位论文

【摘要】：随着现代社会迅速发展,城市化建设日新月异,给城市交通带来了巨大压力,为了降低交通事故发生率,保持交通畅通持续有效运行,必须采取有效的管理方式。传统人工方式耗费大量财力物力而且效率较低,智能交通系统成为未来交通发展的必然选择。在智能交通系统中,行人检测具有重要的实际意义,检测技术可以运用到智能车辆中,有效减少交通事故发生。运用到监控系统中,可以对行人行为进行分析。行人检测算法层出不穷,可以分为两大类:一种是基于模板匹配的传统算法。另一种是基于图像特征提取的检测算法。本文主要研究的是基于图像特征提取的算法。通过对行人检测算法的研究,有效提高检测系统的检测率并降低检测时间。本文首先研究了现有行人检测算法,对现有算法进行了详尽分析,对各种算法的优缺点进行深入比较。现有算法可划分三类:基于全局的检测、基于多部位的检测以及多视角下的检测,本文的研究工作重点主要在前面两类。为了解决行人外貌特征多样性,行人运动规则随意性,选取单一特征最强大的HOG特征,把HOG梯度直方图同SVM支持向量机结合起来,作为行人检测算法,实验结果表明,基于HOG+SVM算法有优秀的识别率。在对分类器训练时,传统的HOG特征提取维数较高,有大量的冗余信息,使得算法计算较为复杂。为了克服这一不足,提出一种改进算法,引入了PCA主成分析法,形成全新的PCA-HOG特征。经过PCA算法处理后能有效减少HOG特征中信息的重叠性,降低特征空间的维度。通过对比单一HOG特征实验结果,PCA算法处理后的分类器训练和检测时间都大大缩短了,同时检测精度也得到了提高。前面分析的HOG+SVM算法是基于全局整体检测算法,在行人未被遮挡下,有优秀的检测效果。在检测过程中,让检测窗口逐一扫描图片上的信息,提取全部特征值来判断有无行人,这种整体检测算法在有遮挡情况下检测效果不理想。为了在行人有被遮挡情况下,提高检测率和检测时间,提出了HOG和LBP多特征融合。这种特征既有行人边缘梯度信息,又有纹理特征信息,能有效弥补单一HOG特征不足之处。由于HOG特征提取计算速度慢,提出一种快速的检测算法,即积分图方法来对HOG特征提取,提高了检测速度。在对分类器设计上设计了两个分类器,一个全身分类器和一个半身分类器,对行人检测时,先送入全身分类器中进行判断,如果判断为无,再送入半身分类器中判断行人是无还是被遮挡了。实验结果表明,在遮挡情况下,该算法有优秀的检测效果。基于前面理论分析,提出一种基于PCA降维的HOG特征,同时融合LBP特征的检测方法,对比实验结果,该算法能有效提高识别率,同时提高训练和检测速度。
[Abstract]:With the rapid development of modern society and the rapid development of urbanization, great pressure has been brought to urban traffic. In order to reduce the incidence of traffic accidents and keep traffic smooth and effective, It is necessary to adopt effective management mode. Traditional manual way consumes a lot of financial and material resources and has low efficiency. Intelligent Transportation system (its) becomes the inevitable choice of traffic development in the future. In the Intelligent Transportation system, pedestrian detection has important practical significance. Detection technology can be used in intelligent vehicles to effectively reduce traffic accidents. In monitoring systems, pedestrian behavior can be analyzed. Pedestrian detection algorithms emerge in endlessly. It can be divided into two categories: one is the traditional algorithm based on template matching, the other is the detection algorithm based on image feature extraction. The detection rate of the detection system is improved effectively and the detection time is reduced. Firstly, the existing pedestrian detection algorithms are studied, and the existing algorithms are analyzed in detail. The existing algorithms can be divided into three categories: global detection, multi-position detection and multi-angle detection. In order to solve the diversity of pedestrian appearance and the randomness of pedestrian movement rules, the most powerful HOG feature with single feature is selected, and the HOG gradient histogram is combined with SVM support vector machine. As a pedestrian detection algorithm, the experimental results show that the algorithm based on HOG SVM has excellent recognition rate. When classifier is trained, the traditional HOG feature extraction dimension is high, and there is a lot of redundant information. In order to overcome this shortcoming, an improved algorithm is proposed, in which the PCA principal component analysis method is introduced to form a new PCA-HOG feature. After the PCA algorithm is processed, the overlap of information in the HOG feature can be effectively reduced. Reduce the dimension of feature space. Compared with the single HOG feature experiment results, the classifier training and detection time are greatly shortened. At the same time, the accuracy of the detection is improved. The HOG SVM algorithm, which is based on the global detection algorithm, has excellent detection effect when the pedestrian is not occluded. In the detection process, let the detection window scan the information on the pictures one by one. In order to improve the detection rate and detection time in the case of pedestrian occlusion, the whole detection algorithm is not effective in the case of occlusion by extracting all the eigenvalues to judge whether there are pedestrians or not. A multi-feature fusion of HOG and LBP is proposed, which has both pedestrian edge gradient information and texture feature information, which can effectively compensate for the shortcomings of a single HOG feature. Due to the slow computation speed of HOG feature extraction, a fast detection algorithm is proposed. In the design of classifier, two classifiers, a full-body classifier and a half-body classifier, are designed. When detecting pedestrians, they are sent to the whole body classifier to judge. If it is judged as no, then it is put into the half-body classifier to determine whether the pedestrian is without or has been occluded. The experimental results show that the algorithm has excellent detection effect in the case of occlusion. A HOG feature detection method based on PCA dimensionality reduction and fusion of LBP features is proposed. Compared with the experimental results, the algorithm can effectively improve the recognition rate and improve the training and detection speed.
【学位授予单位】：成都理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：U495;TP391.41

【相似文献】