智能监视中目标检测及跟踪算法研究

发布时间：2018-03-07 18:53

本文选题：智能视频监控　切入点：运动目标检测　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着人们对于公共安全等问题的不断重视,视频监控被应用到越来越多的场景中,从而带来了海量的监控视频数据。传统人工处理的方式已逐渐不可行,我们迫切需要使用计算机来进行无人值守的智能视频监控。智能视频监控中涉及的关键技术有目标检测及目标跟踪,目标检测根据处理数据对象的不同又可以进一步分为运动目标检测以及目标检测。然而现实场景中由于噪音、光照变化等因素导致运动目标检测算法往往效果不佳;另一方面大部分的检测算法仍面临计算复杂度较高的问题;目标跟踪算法也面临着准确度有限、实时性不强等问题,我们需要寻找更好的解决方案。因此,对于智能监控领域中目标检测及跟踪算法的研究有着重要的理论意义与实用价值。在这种背景下,本文主要进行了以下工作:1.首先介绍了一些经典的运动目标检测算法,包括背景差分法、帧间差分法、光流法以及背景建模法,并分析了各种方法的优缺点。2.重点研究了混合高斯模型(Gaussian Mixture Model,GMM)。GMM是一种很有效的背景建模模型,但是作为一种像素级建模算法,其检测结果存在"空洞"问题,本文在分析GMM以及深度卷积神经网络(Deep Convolutional Neural Network,DCNN)的基础上,提出了一种基于深度编解码网络的运动目标检测算法,实验结果显示我们的算法不仅有效地解决了原GMM算法中的"空洞"问题,而且大幅度提高了算法的鲁棒性。3.介绍了目标检测算法的一般流程,总结并分析了该类算法的关键点。面对当前基于深度学习的目标检测算法中计算复杂度较高的问题,我们引入了一种精确度较高且检测速度很快的目标检测模型Faster-RCNN(Regions with CNN features)。4.研究了现有的一些目标跟踪算法,本文在分析了通用目标跟踪的算法GO-TURN(Generic Object Tracking Using Regression Networks)后使用感兴趣区域池化(Region ofInterest Pooling,ROIPooling)对其改进提出了一种深度回归网络模型,有效地解决了目标跟踪问题,而且该模型很容易集成到现有的网络模型中。实验表明,本文所研究的方法可以很好的应用智能视频监控系统中,并且这些方法可以提高系统的整体性能与效率。
[Abstract]:As people pay more and more attention to public security and other issues, video surveillance has been applied to more and more scenes, which has brought huge amounts of surveillance video data. There is an urgent need to use computers for unattended intelligent video surveillance. The key technologies involved in intelligent video surveillance include target detection and target tracking. The target detection can be further divided into moving target detection and target detection according to the different processing data objects. However, in the real scene, due to noise, illumination and other factors, the moving target detection algorithm is often ineffective. On the other hand, most of the detection algorithms still face the problem of high computational complexity, and the target tracking algorithm also faces the problems of limited accuracy and weak real-time, so we need to find a better solution. It has important theoretical significance and practical value for the research of target detection and tracking algorithms in the field of intelligent surveillance. Under this background, this paper mainly carries out the following work: 1. Firstly, some classical moving target detection algorithms are introduced. It includes background difference method, inter-frame difference method, optical flow method and background modeling method. The advantages and disadvantages of these methods are analyzed. However, as a pixel level modeling algorithm, there is a "hole" problem in the detection results. Based on the analysis of GMM and Deep Convolutional Neural Network (DCNN), a moving target detection algorithm based on deep codec network is proposed in this paper. The experimental results show that our algorithm not only effectively solves the "hole" problem in the original GMM algorithm, but also greatly improves the robustness of the algorithm. 3. The general flow of target detection algorithm is introduced. The key points of this kind of algorithm are summarized and analyzed. Facing the problem of high computational complexity in the current target detection algorithm based on depth learning, We introduce a target detection model Faster-RCNN(Regions with CNN features.4.We study some existing target tracking algorithms. After analyzing GO-TURN(Generic Object Tracking Using Regression networks, a depth regression network model is proposed to solve the problem of target tracking by using region of interest region ofInterest Positioning. And the model can be easily integrated into the existing network models. Experiments show that the methods studied in this paper can be applied to intelligent video surveillance system very well, and these methods can improve the overall performance and efficiency of the system.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TN948.6

【相似文献】