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/