Research On Human Action Recognition

In recent years,automated human action recognition widely applied on video surveillance,computer interaction,virtual reality,etc.has attracted increasing attention in the last decade.Performance on lab captured datasets with controlled environment conditions has been reported to be saturated,thus research efforts has been shifting to multi-view human action analysis under the real world.Despite that,since the introduction of Surveillance Event Detection(SED)task in TRECVID 2008 organized by National Institute of Standards and Technology,it remains a difficult problem and progress of performance improvement is slow.In order to study the above problems,this paper first introduces the existing datasets used in high frequency in the field of human motion recognition,and review the traditional hand-craft features and the emerging deep learning methods in human action recognition respectively.Secondly,the two main research contexts are discussed in detail: One is cross-view human action recognition algorithms.In this paper,the existing cross-view human motion recognition algorithms are used to fuse the feature information of multiple views,so that the model can effectively learn common information of multiple views,which better helps identify human movements.Another one is cross-domain motion recognition algorithms.Cross-domain learning methods are mainly applied to learning from different data distributions.In this paper,the existing cross-domain human motion recognition algorithms are used to solve the multiview in formation extraction problem,and the multi-view common information is applied to represent similar or completely different unseen views in order to better recognize human motion under unseen views.Finally,this paper introduces two different types of multi-view datasets,one dataset is shoot by surveillance cameras to study action recognition from an unseen surveillance view,and the other large-scale one is captured by limited experimental environment conditions used for research the impact of large-scale data on the performance of action recognition.A new database consisting of four datasets obtained under limited experimental environment conditions is used for action recognition and evaluation from unseen view.The above databases are evaluated by specific experimental settings,and the feasibility of the proposed methods is verified through detailed analysis and comparison.