【作者】 陈曦；

【导师】 徐贵力；

【作者基本信息】 南京航空航天大学 ， 工程硕士（专业学位）， 2016， 硕士

【摘要】 大型工地人员车辆情况复杂、安全隐患较多,需要加以有效监督和管理。本文从大型工地的智能监管工程需求入手,研究了面向大型工地的视觉监管系统关键技术。首先,本文针对现有大型工地监管方法主要依赖于人工,存在范围局限、易疏漏、耗费人力物力等缺点的问题,根据工程应用需求和技术指标,设计了面向大型工地的视觉监管系统。其次,针对现有运动目标检测算法受大型工地出入口车辙干扰的问题,在分析像素值历史变化特性的基础上,提出了基于帧间累计变化矩阵的背景差分改进算法,即根据车辙历史帧间累计变化明显弱于当前运动目标的特性,将历史帧间累计变化矩阵中值为零的点更新为背景再进行背景差分的方法。对比实验表明,本文算法能够较好的抑制车辙干扰,与传统背景差、帧间差、帧间差结合背景差的方法相比,前景检测F1指标分别提高了27.3%、14.5%和24.4%。此外,针对基于背景差分的运动目标检测算法原理固有缺陷导致的部分前景分裂问题,从分裂伪子目标的空间位置关系特点入手,设计了基于前景空间信息的运动目标修正算法,能够较好的解决前景分裂问题。然后,针对大型工地对出入人员及车辆的分类统计监管需求,研究了出入人员及车辆情况监管算法。针对应用现场跟踪需求和实时性要求,设计了基于邻帧目标空间关系代价函数的区域跟踪方法;依据目标的运动特性,设计了基于跟踪和计数检测线的目标分类计数统计算法,实现系统对人员及车辆目标的监管功能;根据人员目标分类及施工现场安全要求,针对现有安全帽检测算法易受亮度影响和识别安全帽颜色可靠性有待提高的问题,提出了基于局部HSV颜色模型的安全帽检测算法,即通过人员目标局部HSV颜色特征结合其轮廓特征进行安全帽检测的方法,对比实验表明,安全帽检测正确率达到92.67%,与已有算法相比提高了8.67%。综合实验表明,本文设计的人员及车辆监管算法分类统计准确率达87.18%以上,基本满足大型工地监管要求。最后,根据大型工地现场背景环境突变引起的少量路面伪目标出现问题,在分析伪目标与真实目标的特性基础上,研究了其灰度直方图特征,并研究了屏蔽路面外非前景点的优化方法。对比实验表明,本文提出的灰度直方图丰富度特征对路面伪目标的特征表征效果优于已有的六个灰度直方图特征,综合分类率达到97.3%,能够对路面伪目标进行有效检测。更多还原

【Abstract】 The large-scale construction site needs to take effective management because of the heavy workloads of workers and vehicles, and the potential security risk. This thesis focuses on the implementary requirements of this management and targets the key technology of visual supervision system for the large-scale construction site.Firstly, in order to solve the disadvantage of manual administration such as the limited supervision scope, the latent negligence and the additional resources consuming and so on. A visual supervision system which is designed based on the engineering application requirements is developed for the large-scale construction site.Secondly, aiming at the problem that the existing detection algorithm of the moving target influenced by the wheel trace at the large-scale construction site entrance, an improved background subtraction algorithm based on frame accumulated change matrix is proposed by analyzing the historical change of pixel value in this thesis. The main ideals of proposed method is to update the zero value points of the accumulated change matrix in a frame into the background points according the characteristic that the historical accumulated change of wheel trace region is much weaker than the dynamic moving targets. The experiment results show that the proposed algorithm contributes significantly to decrease the disturbance of wheel trace, and the F1 indicator of the foreground detection is increased about 24.4%, 14.5% and 27.3% compared to the traditional background subtraction method, traditional inter frame difference method and background subtraction method with the inter frame difference. In addition, in order to address the partial foreground splitting problem, which is caused by the inherent defect of the background subtraction algorithm, a moving target correction algorithm based on the spatial information is developed by using the position relationship of the pseudo sub-targets, the experiment shows that can solving this problem of the foreground splitting.Thirdly, according to the supervision demand forlarge-scale construction sitesincluding the classification and statistics of the workers and vehicles in and out, this paper designedan algorithm for supervision of workers and vehicles. At first, according to the application of tracking and real-time requirements, this paper designed acost function based on inter frame target spatial relationship realizing tracking. And then according to the moving characteristics of the target, designed a target classification and statistics algorithm based on tracking and counting line, to realize the classification and counting function of the system for the worker and vehicle targets. What’s more, according to the classification and safety requirements oflarge-scale construction sites, aiming at the problem that existing helmet detection algorithms using RGB model are sensitive to brightness, this paper proposed a helmet detection algorithm based on the local HSV model, which detecting the helmet by the HSV feature and contour feature from a specific region of the worker target, and the accuracy of the experiment is 92.67%, better than the existing algorithm effect by 8.67%, also more suitable for the environment in this paper. Generally, this paper took a comprehensive experiment on the supervision algorithm,and the result shows that the designed algorithm for supervision of workers and vehiclescan classify and count the targets effectively, and its accuracy up to87.18%, basically meet the supervision requirements for large-scale construction sites.Lastly, according to the phenomenonof appearing a small amount of road false targetscaused by background environment mutation in large-scale construction sites, based on the analysis of the gray histogram characteristics of the road false targets and the real targets, aiming at the problem that the existing features cannot distinguish the road pseudo targets out, this paper proposed and constructed the richness feature, and optimizing the feature calculation by removing the interference of the outside road non-front points.The comparative experiment shows that the proposed and constructed richness feature can describing the road false targets better than the existing six gray histogram features, and the comprehensive classification rate up to 97.3%, which can effectively detect the road false targets.更多还原