监控视频中的行人检测技术及其应用

发布时间：2018-05-24 13:00

本文选题：智能监控 + 行人检测　；参考：《国防科学技术大学》2014年硕士论文

【摘要】：传统的视频监控系统仅具备视频的采集和存储功能,过分依赖摄像头硬件,而浪费了大量的前端计算资源,并且非结构化视频数据的监视工作也需要大量人工成本。行人作为事故发生的最主要诱因,是监控视频数据中最重要的组成部分之一,视频中的行人检测技术对监控系统的智能化具有重要意义。本文提出了一种监控视频中的快速行人检测方法,主要从行人检测基本方法、检测分类器的离线训练以及视频中基于运动先验信息的行人检测三个方面进行研究,并基于该方法提出其在智能监控系统的一种应用,即监控视频中行人信息的结构化存储与检索。本文的主要研究内容包括以下几个方面:(1)综合分析了目前主流的几种行人检测方法,最终选取基于积分通道特征和级联Ada Boost分类器的方法作为视频中行人检测的算法基础,并通过大量实验进行分析,评估了该方法的检测性能和可行性,确定了该方法的适用环境;(2)针对Ada Boost分类器的特性和负样本训练集的多样性,提出了一种基于权重退化控制负样本采样的训练方法,加速了行人检测分类器的离线训练过程,并提升了分类器离线训练的精度,为视频中的行人检测打下了坚实的基础;(3)对传统Vi Be方法进行了改进,并基于该方法进行运动检测。根据视频中的运动先验信息,限定积分通道特征的提取,并对视频帧图像分块进行行人检测,有效减小了复杂背景信息的干扰,提升行人目标提取速度的同时,也增强了检测的精度;(4)提出了行人检测技术在智能监控系统中的应用,设计了监控视频中行人检测、结构化存储与行人检索的整体框架,对该项智能监控技术应用进行了探索性研究。
[Abstract]:The traditional video surveillance system only has the function of video capture and storage, so it relies on the hardware of camera too much, and it wastes a lot of front-end computing resources, and the monitoring of unstructured video data also needs a lot of labor cost. As the main cause of accidents, pedestrian is one of the most important components of video surveillance data. Pedestrian detection technology in video is of great significance to intelligent monitoring system. In this paper, a fast pedestrian detection method in surveillance video is proposed, which is mainly studied from three aspects: the basic method of pedestrian detection, the off-line training of detection classifier and the pedestrian detection based on prior motion information in video. Based on this method, an application of this method in intelligent monitoring system is proposed, that is, structured storage and retrieval of pedestrian information in surveillance video. The main research contents of this paper include the following aspects: 1) synthetically analyzing several popular pedestrian detection methods, finally selecting the method based on integral channel feature and cascaded Ada Boost classifier as the basis of pedestrian detection algorithm in video. Through a large number of experiments, the detection performance and feasibility of the method are evaluated, and the suitable environment for this method is determined. (2) aiming at the characteristics of the Ada Boost classifier and the diversity of the negative sample training set, A training method based on weight degradation control negative sample sampling is proposed, which accelerates the off-line training process of pedestrian detection classifier and improves the accuracy of off-line training. A solid foundation is laid for pedestrian detection in video. (3) the traditional Vi be method is improved, and the motion detection is carried out based on this method. According to the prior information of motion in video, the feature extraction of integral channel is limited, and the pedestrian detection of video frame image is carried out, which effectively reduces the interference of complex background information and improves the speed of pedestrian target extraction at the same time. The application of pedestrian detection technology in intelligent monitoring system is proposed, and the overall framework of pedestrian detection, structured storage and pedestrian retrieval in surveillance video is designed. The application of this intelligent monitoring technology is studied in this paper.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41;TN948.6

【参考文献】