Research On Human Action Recognition In Videos

Human action recognition is receiving increasing attention and becoming a hot area ofstudy in computer vision, where recognition of human actions from videos remains to be achallenging work. In recent years, a wide range of promising applications based on humanaction recognition has been introduced. E.g. Advanced human-computer interaction, smartvideo surveillance, 3D TV, movies and animation and gesture-based interactive games, etc.This thesis mainly focuses on recognition of natural human action from videos. Aframework for human action recognition is proposed where human action recognition isseparated into two sub-steps, namely human action presentation and human actionclassification.We conducted human action recognition using Bag-of-Features and video chipping asfollows. First, determine interest points using corner detectors, then analysis the cuboidsaround interest points by calculating their descriptors. For each video, we extract the samenumber of interest points which leads to the same number of descriptors. Then, theBag-of-Features Model is proposed to handle the issue of human action annotation problemwhere K-means algorithm is used for establishing feature dictionary. Video chipping is alsoused to segment the original videos into chips where the annotation of original videos isbased on the human action presented by chips.Recognizing human action based on position distribution of interest points from videosis introduced in this thesis where Harris-OF interest points detection algorithm is adopted andHo P descriptor is proposed. First, determine Harris-OF interest points based on cornerdetectors and optical flow, then calculate Ho P descriptors to describe the position distributionof Harris-OF interest points which being used as descriptor of videos. SVM and NN classifierare used for human action classification. Method based on Ho P avoids mass computationcompared with Bag-of-Features method which makes it incredibly fast.When it comes to human action classification, we adopted segmentation method wherevideos are cut into chips to handle the insufficient memory and training set issue. This thesisalso introduced a new algorithm to improved current chipping method which is calledSmart-Chipping. Experiments show that a reasonable video-chipping not only brings up moretraining samples but improves classification accuracy as well.Bag-of-Features based method was first applied on the Weizmann dataset withstate-of-art results. Ho P based method is capable to recognize human actions from Weizmann,KTH and You Tube datasets with better accuracy than most singly used descriptors, whilethere is a big gap in time efficiency between other methods and Ho P based method.