基于卷积神经网络的密集场景人数估计算法研究

发布时间：2018-02-22 04:45

本文关键词： 深度学习卷积神经网络人群密度估计人群数量估计特征提取　出处：《安徽大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,由于人口的快速增长,因人群密集引起的骚乱事件频繁发生,所以对人群聚集的场合进行监控变得格外重要。采用计算机视觉的方法来对密集场合中的人群进行密度估计、行为分析也成为了当前一个热门研究方向,具有广泛的实际应用价值。本论文针对两方面开展研究:人群区域密度等级估计和人群数量估计,其中密度等级估计是将人群区域划分为如高密集度、中等密集度、低密集度的等级,并在高密集度时进行报警。传统的研究方法大都基于如下框架:1)提取人群区域内的特征,如纹理特征,小波特征。2)通过目标检测或回归模型的方式估计人数。然而在密集的场景中,个体在人群中的分布呈多样性,加之复杂环境下的目标遮挡,光线改变,给高性能的特征描述和人群数量估计带来困难。近年来深度学习算法在计算机视觉领域取得了不断的成功,相关研究表明深度神经网络学习的图像特征较传统图像特征有更强的泛化性、代表性。借鉴前人研究工作的基础上,本文通过大量的试验研究,提出了一种基于深度卷积网络的密集人数估计和人群密度等级分类算法,通过使用两路监督信号的学习方法,增加了网络学习的鲁棒性,抑制了学习过程中可能出现的过拟合。本文的主要工作如下:1)采用卷积神经网络来对图像区域内的人群特征进行提取。经过对网络特征的学习将区域内的人群划分密度等级从而完成人群高密度预警的任务。2)提出了两路监督信号的学习方法。一路信号用来分类区域内的人群密度等级,另一路信号用来对区域内的人群数量进行回归学习,实现对区域内的人群数量估计。两路信号的学习方法也增加了网络学习的鲁棒性,提高了人数估计结果准确性。3)在实验验证中,我们在自己创建的数据集AHU-CROWD,以及当前常用的公开数据集,包括UCF-CROWD,UCSD上进行了实验。结果表明相比于之前的研究,本论文方法在人数估计准确率上有了显著的提高。
[Abstract]:In recent years, due to the rapid growth of the population, riots caused by dense crowds have occurred frequently. Therefore, it is particularly important to monitor the situation where people gather. Using computer vision to estimate the density of people in dense situations, behavior analysis has also become a hot research direction. This paper focuses on two aspects: the estimation of population regional density grade and the estimation of population quantity, in which the population region is divided into two groups, such as high density, moderate density, and population density. Traditional research methods are based on the following framework: 1) to extract features in crowd areas, such as texture features. Wavelet feature. 2) estimate the number of people by target detection or regression model. However, in the dense scene, the distribution of individuals in the crowd is diverse, combined with the target occlusion in complex environment, the light changes. In recent years, deep learning algorithms have been successful in the field of computer vision. Related studies show that the image features of depth neural network learning have stronger generalization and representativeness than traditional image features. In this paper, a classification algorithm for dense population estimation and population density classification based on deep convolution network is proposed. By using two supervised signals, the robustness of network learning is improved. The main work of this paper is as follows: 1) Convolutional neural network is used to extract the crowd characteristics in the image region. The study method of two supervised signals is put forward. The first signal is used to classify the population density grade in the region. The other signal is used to learn the number of people in the region by regression, and to estimate the number of people in the region. The learning method of two signals also increases the robustness of network learning. In experimental verification, we conducted experiments on the data set AHU-CROWDcreated by ourselves, as well as on currently commonly used open datasets, including UCF-CROWDU UCSD. The results show that compared with previous studies, In this paper, the accuracy of the estimation of the number of people has been improved significantly.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】