智能视频监控中行人的检测与跟踪方法研究

发布时间：2018-01-12 10:43

本文关键词：智能视频监控中行人的检测与跟踪方法研究　出处：《北京交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：智能视频监控在获取监控视频数据基础上,对场景中的目标如车辆,人进行检测,检测的方法是利用目标的一些运动特征或者外观特征如颜色,纹理等结合检测窗口,检测窗口可以是基于感兴趣区域或者显著性区域,也可以利用滑动窗口遍历。在检测基础上可以进一步跟踪目标,获取目标在一段视频序列内的轨迹。目标的检测和跟踪是下一步目标动作识别,行为分析的基础,但是由于场景中的背景时刻在变化且有光照,噪声的影响再加上目标之间相互遮挡给目标的检测和跟踪增加了难度。同时以往的行人检测方法主要是在公有数据集的基础上提取特征然后利用分类器模型训练,所得行人检测器在原始数据集上往往能得到较高的准确率。但是一旦应用到其他场景中,检测率将大大下降。本文提出了一种基于迁移学习和稀疏编码的行人检测框架,该框架可以将在原训练集上训练好的行人检测器迁移到新场景中,该框架中首先将原始检测器应用到目标场景中获得初始检测结果,然后利用一些线索过滤出那些被检测器正确分类的样本作为目标模板,然后利用稀疏编码刻画目标模板和目标样本之间的相似性并且加权目标样本。同时,利用显著性检测方法检测目标模板和原训练集中行人样本的显著性区域,并利用稀疏编码加权原训练集中的样本,最后利用支持向量机训练所有带权值样本得到目标场景下的行人检测器。基于迁移学习所得的该检测器在特定新场景中检测率比原始检测器提高了近30%。本文同时还提出一种基于粒子滤波行人跟踪的框架,详尽的阐述了粒子滤波框架的原理,即如何从贝叶斯理论和蒙特卡洛过渡到粒子滤波方法理论。基于行人颜色特征和粒子滤波框架下结合上一步检测方法的验证实现新场景下行人的跟踪且实验证明该跟踪方法较传统的跟踪方法具有更好的鲁棒性。
[Abstract]:Intelligent video surveillance is based on the acquisition of video data, the scene of the targets such as vehicles, people to detect, the method of detection is to use some of the moving features of the target or appearance features such as color. Texture and other combined detection window, detection window can be based on the region of interest or significant region, or can be traversed by sliding window, on the basis of detection can be further tracking the target. Target detection and tracking is the basis of the next target action recognition and behavior analysis, but because the background of the scene is changing and there is light. The influence of noise and the mutual occlusion between targets increase the difficulty of target detection and tracking. Meanwhile, the previous pedestrian detection methods are mainly based on the common data set to extract features and then train by classifier model. . The obtained pedestrian detector can obtain high accuracy in the original data set, but once applied to other scenarios. The detection rate will be greatly reduced. This paper proposes a pedestrian detection framework based on migration learning and sparse coding, which can transfer the trained pedestrian detectors on the original training set to a new scene. In this framework, the original detector is first applied to the target scene to obtain the initial detection results, and then some clues are used to filter out the samples correctly classified by the detector as the target template. Then sparse coding is used to describe the similarity between target template and target sample and weighted target sample. At the same time, significance detection method is used to detect significant area between target template and pedestrian sample in the original training set. The sample of the original training set is weighted by sparse coding. Finally, support vector machine (SVM) is used to train all weighted samples to obtain pedestrian detectors in target scenarios. The detection rate of the detector based on migration learning is 30% higher than that of the original detector. At the same time, a framework of pedestrian tracking based on particle filter is proposed. The principle of particle filter framework is described in detail. That is, how to transition from Bayesian theory and Monte Carlo to particle filter theory. Based on pedestrian color characteristics and particle filter framework combined with the verification of the previous detection method to achieve the new scene downlink tracking and experimental results. This tracking method is more robust than the traditional tracking method.
【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】：TN948.6

【参考文献】