摘要: This thesis aims to present novel description methods for human action recognition.Generally，a video sequence can be represented as a collection of spatial temporal words by detecting space-time interest points and describing the... 展开 This thesis aims to present novel description methods for human action recognition.Generally，a video sequence can be represented as a collection of spatial temporal words by detecting space-time interest points and describing the unique features around the detected points (Bag of Words representation).Interest points as well as the cuboids around them are considered informative for feature description in terms of both the structural distribution of interest points and the information content inside the cuboids.Our proposed description approaches are based on this idea and making the feature descriptors more discriminative.In this thesis，we propose and validate three types of description methods：the first is projected 3D Shape Context (projected 3DSC) which is derived from the original shape context and makes use of the structural distribution of interest points; the second are the Transform based description methods which are widely exploited in image processing and utilizes the appearance information of images; the third is Correlogram of Oriented Gradient (COG) which is built from the spatial temporal gradients of each cuboid taking advantage of both the spatial structure and appearance information.The proposed methods are tested in the very challenging and well studied KTH dataset.The projected 3DSC improves the classification accuracy by 10％ compared to that of the original 3DSC.And the Wavelet Transform based descriptor achieves as high as 93.89％ recognition rate，which is better than any of the state-of-the-art methods.Correlogram of Oriented Gradient achieves 15％ better than the popular Histogram of Oriented Gradient (HOG).Further more，we validate the efficiency of the Wavelet Transform based method on a more realistic and challenging human action dataset：the Hollywood dataset. 收起