室内监控中移动检测与跟踪算法的改进与实现

发布时间：2018-01-07 04:05

本文关键词：室内监控中移动检测与跟踪算法的改进与实现　出处：《东南大学》2017年硕士论文　论文类型：学位论文

【摘要】：基于智能视频监控的移动目标检测、识别与跟踪是计算机视觉领域研究的热点,在现代室内安全防护系统中有着越来越多的应用,使用这些技术,我们可以快速获取监控区域中感兴趣的前景目标、识别前景目标,并对前景目标跟踪形成跟踪轨迹,为后续目标的行为分析与理解打下良好的基础。本文以室内监控环境为研究场景,以单目标人体为研究对象,并将单目标人体的检测、识别以及跟踪作为本文研究的主要内容,旨在通过对比分析现有的移动目标检测与跟踪算法,改进现有算法的某些不足,避免监控过程中常见的干扰,以提升室内智能监控系统的鲁棒性。本文的主要工作如下:首先,在移动目标检测阶段,针对背景减差算法对光线变化比较敏感的缺点,本文提出了基于GMM算法的背景减差算法,利用GMM算法良好的稳定性以及对光线缓慢变化不敏感的特点,为静态背景图像建立背景模型。另外,针对GMM算法对光照突变适应性差的缺陷,则通过定义前景目标所占的面积比率以及光线突变持续的帧数来检测室内光线是否发生突变。实验表明,改进的移动目标检测算法不仅可以完整检测出前景目标,而且对于传统的背景减差算法光线缓慢变化以及突变情况引起的检测误差也可以很好地解决,从而大大提升了移动目标检测的准确率以及查全率。其次,在人体目标识别阶段,针对室内监控环境下,不同前景目标的分类问题,本文提出基于HOG特征的SVM分类器算法对前景目标进行分类,通过借助公共数据集INRIA提供的正负样本进行分类器训练。最后通过仿真实验验证了该分类算法具有较高的准确率。最后,在单目标人体跟踪阶段,针对传统Camshift移动目标跟踪算法抗遮挡性差以及目标尺度变化过大敏感性的缺点,本文提出了一种改进的Camshift目标跟踪算法。采用对目标分块跟踪的方式来处理目标遮挡的问题,并通过定义目标匹配率来判断目标不同程度的遮挡。另外,针对目标尺度变化过大引入的跟踪误差,本文通过将目标的几何特征和目标的颜色特征结合起来,以更充分地描述目标,提高目标的识别率。实验表明,改进的移动目标跟踪算法在保证系统实时性的前提下,对于传统的Camshift跟踪算法抗遮挡性差以及尺度变化过大带来的跟踪误差都能很好地解决,提高了移动目标跟踪阶段的鲁棒性。
[Abstract]:Moving target detection based on intelligent video surveillance, recognition and tracking is a hot research field of computer vision, it has more and more application in modern interior safety protection system, the use of these techniques, we can quickly get the foreground object interested in the monitoring area, identify the foreground objects, and the prospect of the target tracking tracking trajectory lay a good foundation for the follow-up behavior analysis and understanding of the target. This paper takes the indoor monitoring environment of the scene, with a single target body as the research object, and the detection of single target recognition and tracking of the human body, as the main content of this paper is to, through the comparative analysis of moving target detection and tracking algorithm in the existing, some overcome the shortcomings of the existing algorithms, avoid common interference in the process of monitoring and control, to enhance the robustness of indoor intelligent monitoring system. The main work of this paper is as follows: first of all In moving target detection, background subtraction algorithm, aiming at the light sensitive shortcomings, proposed subtraction algorithm GMM algorithm based on the background, using the GMM algorithm and good stability to light slow change characteristics is not sensitive to the static background images to establish the background model. In addition, according to the light mutation low adaptability of the GMM algorithm, through the definition of the area occupied by the foreground object and the ratio of the number of frames to detect the light mutation for interior light mutation. Experimental results show that the moving target detection algorithm can not only detect foreground objects, but also for the traditional background subtraction detection error caused by poor light slow change and mutation algorithm the situation can be solved very well, thus greatly enhance the accuracy of moving target detection and recall. Secondly, the human target recognition stage According to the monitoring, indoor environment, classification of different objects, this paper proposes a SVM classification algorithm based on HOG feature of foreground object classification, classifier training by means of positive and negative samples of public data sets provided by INRIA. The simulation experiment verifies the accuracy of the algorithm has a high classification. Finally, in the single target tracking the body, according to the traditional Camshift mobile target tracking algorithm for anti block difference and the scale change of target large sensitivity tracking algorithm is proposed in this paper an improved Camshift target. To deal with object occlusion problem using the target block tracking, and the matching rate to determine the occlusion target different degrees by definition of target. In addition, the tracking error for the target scale change is too large is introduced, this paper will combine the color features and geometric features of the target Up to more fully describe the target, improve the recognition rate. Experimental results show that the algorithm not only guarantees the real-time system under moving target tracking is improved, the traditional Camshift tracking algorithm of anti occlusion and scale variation of the tracking error caused by the large can be a good solution to improve the robustness of mobile the target tracking stage.

【学位授予单位】：东南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】