【作者】 黄炜；

【导师】 黄立勤；

【作者基本信息】 福州大学 ， 通信与信息系统， 2017， 硕士

【摘要】 在大数据时代的当下,随着世界人口的快速增长与技术的进步,基于视频的人群统计方法有着无比广泛的应用,包括商场顾客行为分析,公共场所安全监控,企业管理等。在国家自然科学基金(61471124)和福建省重点科技项目(NGII20160208)的支持下,本课题对人群统计方法的相关理论技术展开研究,并针对中低密度人群场景与动态场景(包含多种密度人群场景),提出两套复杂度不同的解决方案。一、在中低密度场景下,基于检测的人群统计方法虽然复杂度低,通用性强,但在较高密度场景下统计精度不高,而基于回归的方法虽然在不同密度场景中均有不错的统计精度,但因为需要针对特定场景做预处理(视角失真纠正),所以通用性差,又因为其复杂度高,所以用在低密度人群场景性价比低。故在中低密度人群场景下并没有一个复杂度低,通用性强,且性能卓越的人群统计算法。针对该问题,本课题提出了基于可变形部件模型的人群统计方法。该方法使用基于可变形部件模型的行人检测器获得检测候选框,预处理后输入迪迪克雷过程混合模型进行聚类,对于聚类的结果使用改进型基于角点的计数方法,对于获得的预计数值,使用平均滤波器进行平滑,得到最终计数值。该人群统计方法通过聚类将基于检测和基于回归两种人群统计方法融合起来,既有基于检测方法的复杂度低,通用性强的优点,又有基于回归方法在较高密度场景上的性能。实验证明,该方法在中低密度场景相较于其它算法取得了较高的准确率。二、目前算法主要针对高密度人群场景或者中低密度人群场景实现高精度计数,不存在一个能在动态场景(包含高中低密度)下实现高精度计数的算法。针对该问题,本课题提出了基于特征点数量与人群密度相关性的人群统计方法。该方法使用基于累计通道特征的行人检测器进行行人检测,并使用混合高斯模型提取的前景图对检测候选框进行预处理,接下来使用迪迪克雷过程混合模型进行聚类,用改进型基于角点的人群计数算法预计数后,根据本课题提出的基于特征点数量与人群密度相关性的密度估计算法,来获得人群密度,并根据不同密度,使用基于上下文的数据融合方法。实验证明,本课题方法在中低、高密度人群场景上,都能达到较高的统计准确率。本课题针对传统的人群统计方法中存在的问题,提出了两套复杂度不同,适用于不同密度情况的解决方案,实验证明,本课题设计的算法,具有良好的性能以及较高的实用价值。更多还原

【Abstract】 At the age of big data,with the rapid growth of the world population and technology progress,the population statistics method based on video has a very wide application,including the analysis of customer behavior,public safety monitoring,enterprise management etc.Under the support of the National Natural Science Foundation(61471124)and Fujian province key science and technology project(NGII20160208),we study the theory and methods of crowd statistical methods,and put forward two sets of solutions with different complexity for low and medium density crowd scenes and dynamic scenes(including a variety of density crowd scenes).First,in medium-low density scene,the detection based population statistical method is low in complexity and versatility,but the statistical accuracy is not high in the high density scene.The regression based method although have good statistical accuracy in different density scenes,but because of the need to do preprocessing for specific scenarios(perspective distortion correction),so the versatility is poor.Besides,because of its high complexity,it cost too much in low density scene.Therefore,there is not a low complexity,strong universality,and excellent performance of crowd statistical algorithm in the medium-low density scenes.To solve the first problem,the thesis proposes a Deformable part model based crowd counting method.The method uses a deformable part model based pedestrian detector to obtain the candidate windows,and then the processed candidate windows are input into Dirichlet Process Mixture Model(DPMMs)to be clustered.After that,the improved corner points based counting method is used to count each clusters,after that,the average filter is used to smooth the values,and then the final value is obtained.This algorithm fuses the detection based and regression based counting methods by clustering algorithm.It makes this algorithm to be simple and widely using like detection based methods,and to be counting precise in higher density scene like regression based methods.Experimental results show that the proposed method achieves good results in medium-low density scenes.Second,the most of existing algorithms mainly for high density or medium-low density to achieve high accuracy counting,there is no one can achieve high precision counting in dynamic scene(including different density scenes).To solve this problem,this paper puts forward the features number and people density correlation(FNPDC)based crowd counting method.This algorithm uses the aggregate channel features based pedestrian detector to detect pedestrians in the scene,then the foreground map obtained by Gaussian Mixture Models(GMMs)is used to filter the candidate windows.After that,the Dirichlet Process Mixture Model(DPMMs)is used for cluster and the improved corner points based counting method is used for pre-counting.Next,the proposed features number and people density correlation based crowd density estimation method is used to estimate the crowd density,finally this paper make a context aware data fusion based on the estimated crowd density.The experimental results show that this method can achieve the high accuracy in the medium-low and high density crowd scenes.As for solving the problems of traditional crowd counting methods,this paper proposed two algorithms with different complexity for different crowd density.The experiment results show that our algorithms have excellent performance and high practical value.更多还原