Research On Action Recognition Based On Spatio-temporal Features

The main research of this dissertation is the recognition of human actions inrealistic video data, such as movies，video surveillance and some standard databasewhich are used for parameter train and test. There is a large demand for actionrecognition, such as human computer interaction. To this end, we use thespatio-temporal interest point methods to encode the action information from thevideo. We use some usual detector to detect the interest points and descript themaround them by a block of the video. And the experiment proves a good result on theaction classification and location especially on the confuse surround and the realisticvideo. The main contribution of the work of this dissertation is as follows.1. We proposed an improved local descriptor which can be applied on the cuboidsdetector.We know that a common trend in object recognition is to detect and leveragethe use of sparse, informative features points. The use of such features makesthe problem more manageable while providing increased robustness to noiseand pose variation. We detect the spatio-temporal interest points with theGaussian filter on the spatial domain and Gabor filter on the time domain. Andwe proposed an improved descriptor with the surf which can decrease the taskof the complex computation on the video analysis. The experiments shows ouralgorithm perform good result on the recognition.2. We introduces a method which can merge the HOG3D spatio-temporaldescriptor with the detector Harris3D for local features.Building on the success of descriptors based on histograms for orientedgradients(HOG) for static images, we view videos as spatio-temporal volumesand generalize the key HOG concepts to3D. we use the HOG3D descriptor todescript the spatio-temporal interest points and the detector of the points weuse the Harris3D which is proposed by Laptev. Descriptor parameters areevaluated on four action datasets(KTH, Weizmann, Youtube, Hollywood2) andare optimized for action recognition.3. We use the optical flow that can form a local feature representation for videorepresentation for video sequences using the HOG and HOF based on featuretrajectories.And we merge the HOG and HOF features to form the final action featurevector. The HOG features represent the appearance and the HOF represent themotion information. We put them together so that we can get rich informationabout the action in the video. The experiments shows a good performance on HOG and HOF.