• video analysis:crowd counting


       本科毕设:监控视频下的人数统计

       核心点:crowd counting

       目前主要的实现方法: counting by detection,counting by clustering, and counting by regression.

               counting by detection:

                           解决思路:detection+tracking

                           主要问题:detector在复杂高密度人群中性能差;time-consuming

               counting by clustering:

                           解决思路:motion cluster

                           主要问题:要求视频帧率足够支持motion分析;

              counting by regression:

                           解决思路:将低维特征映射到people count

                           主要问题:映射的精度;

           综上:counting by detection是比较直观的做法,但跑了demo之后结果不甚满意;目前我比较中意 counting by regression的解决思路,也在着手阅读相关的论文。

      测试数据:

         目前下载的几个数据库预览:

                      UCSD行人                                    PETS2009                        PETS2002                                  mall

           

          UCSD行人库和PETS2009均为室外库,分辨率较好,直接跑human detector效果也是稳定的;

          PETS2002,mall为室内库,分辨率较差,直接跑human detector(RCNN)效果比较差;【可能下周能从项目上拿到其他的零售店的video】

          现在看的论文很多都是针对特定场景训练的,目前也不能做跨场景的,所以尽快确定场景,才能尽快开始工作。

     state of art:ICCV15(还没有调研完全)

               

         测试标准: mean absolute error (mae), mean squared error (mse), and mean deviation error (mde) 

     相关工作:

                •ICCV15:Bayesian Model Adaptation for Crowd Counts
                • Trans on image processing12:Counting people with low-level features and Bayesian regression
                •ICCV09:Bayesian Poisson regression for crowd counting
                •Cvpr08:Privacy preserving crowd monitoring: Counting people without people models or tracking
                •AVSS12:People Count Estimation In Small Crowds
                •ICCV13:From Semi-Supervised to Transfer Counting of Crowds
                •ICCV15:COUNT Forest: CO-voting Uncertain Number of Targets using Random Forest for Crowd Density Estimation
                •ICIP14:CROWD ANALYSIS IN NON-STATIC CAMERAS USING FEATURE TRACKING AND MULTI-PERSON DENSITY
                •CVPR15:Person Count Localization in Videos from Noisy Foreground and Detections
                •CVPR15:Cross-scene Crowd Counting via Deep Convolutional Neural Networks
                •CVPR13:Cumulative Attribute Space for Age and Crowd Density Estimation
                •BMVC12:Feature Mining for Localised Crowd Counting
     相关研究组:
                https://scholar.google.com/citations?hl=en&user=Fykyo9gAAAAJ&view_op=list_works&sortby=pubdate
                http://personal.ie.cuhk.edu.hk/~ccloy/
               
  • 相关阅读:
    Linux安装.net core
    在IIS中如何配置SSL(https)
    Lambda的前世今生
    Java建造者模式
    leetcode 83 Remove Duplicates from Sorted List
    【设计】Office
    【设计】简书品牌重构
    【设计总结】
    【交互设计】
    【设计总结】
  • 原文地址:https://www.cnblogs.com/xy2012/p/5092206.html
Copyright © 2020-2023  润新知