人群分析综述--Crowd Scene Understanding from Video: A Survey

Crowd Scene Understanding from Video: A Survey ACM Trans. Multimedia Comput. Commun. Appl., Vol. 13, No. 2, Article 19, Publication date: March 2017

针对人群密集场景理解,这篇综述主要关注两个方向:crowd statistics and behavior understanding 人群统计和行为理解

2 Crowd counting 人群计数目前从两个角度来解决,一类方法是回归计数,直接得到图像中总人数的估计,另一类方法是先估计密度图,再有密度图积分求和得到总人数。目前来说密度图提供的信息更多些,得到的精度相对也更高些。这里我们将所有人群计数方法划分为六类:pixel-level analysis, texture-level analysis, object-level analysis, line counting, density mapping, and joint detection and counting

2.1. Pixel-Level Analysis 基于像素水平的分析主要从边缘检测开始,用边缘特征(长度、方向等)来训练一个模型用于人群密度图的估计 begin with edge detection and use edge features (length, orientation, etc.) to train a model

2.2. Texture-Level Analysis 基于纹理的分析对象是图像块,通常使用 statistical, structural, or spectral analyses 来提取特征 coarser grained than pixel-based methods and aims to analyze image patches,use statistical, structural, or spectral analyses to extract features.

基于像素和基于纹理都是得到总人数,它们不能区分出个体

2.3. Object-Level Analysis 检测出单个个体,精度相对较高,但是只适用于低密度场景 identify individual subjects in a scene, applicable to lower density crowds 这里有一篇文献是基于深度和颜色信息来统计人流量,相机安装方式是垂直的。 People-flow counting in complex environments by combining depth and color information

2.4. Line Counting 这类方法主要统计穿越一条线的人流量

2.5. Density Mapping 估计密度图,目前很多基于CNN深度学习的方法都是用于估计人群密度图的 estimates the density of the scene

2.6. Joint Detection and Counting 利用密度图来更好的做检测和计数 makes use of density maps to better detect and count objects within the scene

Challenging problems still exist in measuring large crowds with high densities. These challenges also require good data with images taken at a variety of standoffs, angles, and resolutions and ground-truth labels for comparisons.

人群计数的挑战性还是很大的,尤其是高密度人群的场景,还是非常需要大量数据来驱动算法性能的提升。

3 Behavior analysis 最基本的行为分析就是 normal behavior and abnormal behavior,对于人群密度场所我们主要提取 Dominant motion,对于不是很密度的场景,则是分析 group interactions,也就是小人群的行为。本文主要关注以下三个方面的内容:1)abnormal behavior detection,2)dominant motion extraction,3) group behavior detection

3.1. Abnormal Behavior Detection 什么是异常行为了? any deviation from the normal crowd flow is considered abnormal activity 这里主要基于场景中的运动信息来判断是否有异常行为的存在,前后运动信息的对比

3.2. Dominant Motion Extraction 主要是提取图像中的主要运动信息

3.3. Group Behavior Detection 主要侧重小规模人群聚集的行为分析

3.4. Crowd Analysis and Tracking 对于密集人群的跟踪难度还是很大

4 Crowd video datasets

4.1. Crowd Counting 主要有以下几个数据库:The WorldExpo’10 Crowd Counting Dataset [Zhang et al. 2015], the UCSDdataset [Chan et al. 2008], and the UCF Crowd Counting dataset [Idrees et al. 2013]

4.2. Group Detection 主要有以下几个数据库:The MuseumVisitors [Bartoli et al. 2015], Collective Motion [Zhou et al. 2014], student003 [Solera et al. 2015], and GVEII [Solera et al. 2015] datasets were constructed for group detection. In addition, the Grand Central Station dataset [Zhou et al. 2012] and the Mall dataset [Loy et al. 2013] contain similar attributes as the those previously listed in this group. MuseumVisitors is a dataset for pedestrian and group detection, gaze estimation, and behavior understanding.

4.3. Behavior Understanding The Unusual Crowd Activity dataset [University of Minnesota 2006], 2009 Performance Evaluation of Tracking and Surveillance (PETS2009) dataset [Ferryman and Shahrokni 2009], and Collective Activity dataset [Choi et al. 2009] are purposed for crowdbehaviorunderstanding.

4.4. Holistic Crowd Movement The Meta-Tracking for Video Scene Understanding dataset [Jodoin et al. 2013], Data-Driven Crowd Analysis dataset [Rodriguez et al. 2011b], Crowd Segmentation dataset [Ali and Shah 2007], and the Chinese University of Hong Kong (CUHK) Crowd dataset [Shao et al. 2014] are largely developed from online video repositories

4.5. Synthetic The Agoraset dataset [Allain et al. 2012] is a graphical rendering of crowd sequences with associated ground truths, which include include individual trajectories of pedestrians inside the crowd, crowd density, and the velocity field. Seven environments, also called scenes, were created using crowd models proposed by Helbing et al. [2000]

总体来说深度学习对于人群分析正处于方兴未艾阶段!

本文参与腾讯云自媒体分享计划,欢迎正在阅读的你也加入,一起分享。

发表于

我来说两句

0 条评论
登录 后参与评论

相关文章

来自专栏AI2ML人工智能to机器学习

概率估值的光滑(Smoothing)

一般通过频率(Frequency / Counting)来估值的过程中,存在需要光滑(Smoothing)的问题, 这个问题在自然语言处理或者贝叶斯算法中尤为突...

662
来自专栏GAN&CV

强化学习用于发掘GAN在NLP领域的潜力

本文转载自: https://mp.weixin.qq.com/s?__biz=MzUzNTA1NTQ3NA==&mid=2247484192&idx=1&...

863
来自专栏机器之心

观点 | 下一步研究目标:盘点NLP领域最具潜力的六大方向

选自ruder.io 作者:Sebastian Ruder 机器之心编译 参与:李泽南、黄小天 在开始你的研究之前,了解目标领域中最重要的研究方向是很重要的任务...

2634
来自专栏专知

专知主题链路知识推荐#1——马尔科夫链蒙特卡洛采样(附代码)

【导读】主题链路知识是我们专知的核心功能之一,为用户提供AI领域系统性的知识学习服务,一站式学习人工智能的知识,包含人工智能( 机器学习、自然语言处理、计算机视...

2797
来自专栏技术随笔

【译】用于肺部CT肺结节分类的深度特征学习摘要

3648
来自专栏机器之心

当前训练神经网络最快的方式:AdamW优化算法+超级收敛

Adam 优化器之旅可以说是过山车(roller-coaster)式的。该优化器于 2014 年推出,本质上是一个出于直觉的简单想法:既然我们明确地知道某些参数...

762
来自专栏专知

【业界】 | 谷歌 NIPS 2017 23篇论文:从注意力到价值预测网络(附论文下载)

【导读】2017年度神经信息处理系统大会(NIPS 2017)将于12 月份在美国长滩举行,本届NIPS共收到 3240 篇论文投稿,录用 678 篇,录用率为...

34010
来自专栏量子位

汤晓鸥为CNN搓了一颗大力丸

是重新搭一个模型呢,还是拿来新数据重新调参,在这个已经训练好的模型上搞迁移学习呢?

580
来自专栏机器之心

业界 | 谷歌发布TensorFlow 1.4与TensorFlow Lattice:利用先验知识提升模型准确度

2666
来自专栏CVer

[计算机视觉论文速递] 2018-03-03

通知:这篇推文很长,有32篇论文速递信息,涉及目标检测、图像分割、网络优化、人脸表情识别、SLAM和OCR等方向。 [1]《The 2018 DAVIS Cha...

41412

扫码关注云+社区