面向公共环境的行人头部检测方法研究

发布时间：2018-11-15 16:23

【摘要】：近年来,随着人们对公共安全问题的日益关切,对公共环境中的行人进行自动检测已成为智能监控和计算机视觉领域中的研究热点。由于行人可能处于移动或静止等不同的状态,且姿态变化剧烈,此外,行人之间的移动容易导致相互遮挡,在复杂场景中的这样遮挡情况就更加明显,这些都在很大程度上加大了行人检测的难度。由于相对于人体的其他部位,人体头部不易被遮挡,且轮廓边界比较分明。因此通过检测行人头部来对行人进行定位是一种可行的检测思路。本文对复杂场景下的人体头部检测问题展开研究,致力于研究一类可以对公共环境中人体头部进行稳定检测的算法。论文主要工作内容如下:(1)对固定摄像头情景下的行人头部检测问题进行研究。首先采用背景差分法得到运动前景区域,然后通过分析人体头部轮廓的梯度和方向幅值得到可能存在行人头部的感兴趣区域,同时,采用K-mean聚类方法对伪候选区域进行排除;最后对感兴趣区域训练分类器进行分类得到正确的行人头部位置。实验表明该方法在保持较高准确率的同时,提高了算法的实时性。(2)设计一种摄像头运动情景下的头部检测算法。当摄像头安置于移动终端或车辆等运动物体之上时,背景检测方法变得不再可靠。为此本文采用基于图分割的方法对图像进行分割。通过图分割可以列举出图像中物体可能存在的区域,将这些区域作为头部目标可能存在的区域进行分类检测即可得到目标位置。实验结果验证了方法的有效性。(3)针对传统的PCA-HOG降维后的特征空间存在鉴别性不足的问题,提出一种基于HOG-LBP分步降维的人体头部分类算法。在对传统的HOG检测算法进行分析的基础上,针对PCA降维并没有考虑样本类别的问题,将HOG特征按照样本类别进行分步降维,并将降维后的HOG特征与LBP特征相融合得到最终特征描述算子。通过训练支持向量机分类器进行分类的结果表明,该方法与传统的HOG算子相比可以有效提高检测的准确率。(4)搭建了行人头部检测系统的软件构架。以满足行人头部检测系统的实际需求为目标,在Visual Studio2010+OpenCV框架下开发了相应的检测软件。为了提高软件的稳定性和实时性,在结构上采用多线程技术将图像处理和界面数据交互分为不同的线程独立执行。实际运行结果表明,所开发的软件可以稳定运行。
[Abstract]:In recent years, with the increasing concern of public safety, automatic detection of pedestrians in the public environment has become a research hotspot in the field of intelligent monitoring and computer vision. Since pedestrians may be in different states, such as moving or stationary, and their posture changes dramatically, in addition, the movement of pedestrians can easily lead to mutual occlusion, which is even more obvious in complex scenarios. These have greatly increased the difficulty of pedestrian detection. Compared with other parts of the human body, the head of the human body is difficult to be occluded, and the contour boundary is clear. Therefore, it is a feasible way to detect the pedestrian by detecting the head of the pedestrian. In this paper, the problem of human head detection in complex scenes is studied, and a class of algorithms that can detect human head stably in public environment are studied. The main work of this paper is as follows: (1) the detection of pedestrian head under fixed camera is studied. The background difference method is used to obtain the moving foreground area, and then by analyzing the gradient and directional amplitude of the human head contour, the region of interest that may exist in the pedestrian head is obtained. At the same time, The pseudo candidate regions are excluded by K-mean clustering method. Finally, the correct position of pedestrian head is obtained by classifying the region of interest training classifier. Experiments show that this method not only keeps a high accuracy, but also improves the real-time performance of the algorithm. (2) A head detection algorithm is designed under the camera motion scenario. When the camera is mounted on moving objects such as mobile terminals or vehicles, background detection methods become less reliable. Therefore, the image segmentation method based on graph is adopted in this paper. The possible regions of the objects in the image can be listed by image segmentation, and the location of the objects can be obtained by classifying these regions as the possible regions of the head target. The experimental results demonstrate the effectiveness of the method. (3) aiming at the problem of insufficient discrimination in the traditional PCA-HOG dimensionality reduction feature space, a human head classification algorithm based on HOG-LBP step reduction is proposed. Based on the analysis of the traditional HOG detection algorithm, aiming at the problem of PCA dimensionality reduction without considering the sample category, the HOG feature is reduced according to the sample classification step by step. Finally, the final feature description operator is obtained by merging the reduced HOG features with the LBP features. The results of training support vector machine classifier show that compared with the traditional HOG operator, this method can effectively improve the accuracy of detection. (4) the software architecture of pedestrian head detection system is built. In order to meet the actual needs of pedestrian head detection system, the corresponding detection software is developed under the framework of Visual Studio2010 OpenCV. In order to improve the stability and real-time performance of the software, the multi-thread technique is used to separate the image processing and interface data into different threads. The actual operation results show that the developed software can run stably.
【学位授予单位】：淮北师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】