Research On Human Action Recognition

Human action recognition is very important in the field of computer vision.Human action recognition classified the action from the video or the image by thecomputer. This paper did some research work about the human action recognitionmethods.The human action recognition consists of preprocessing, feature extraction andaction classification. In the preprocessing stage, this paper mainly introduces threemethods of target detection and compares the three methods. In terms of featureextraction, this paper introduces the Gist feature, STIP feature, Dollar feature, Hog3Dfeature, HON4D feature and skeleton feature, and then expounds the extractionprocess in detail. In terms of action classification, this paper first introduces thediscriminative model and generative model. These models include Na ve BayesianClassification Model, Hidden Markov Model, Random Forest Model and SupportVector Machine. Secondly, this paper introduces the Conditional Random Fields fromthe graph structure, learning and inference algorithm.In this paper, there are two major studies. One is under the same body features;the paper designs the experiment to compare the different classification models andanalyzes their advantages and disadvantages. The other one is under the classificationof the random field model; the paper designs experiment to explore the superiorityand inferiority of different body features. In addition, the paper analyzes theshortcomings of current human action recognition datasets and records a new dataset.The paper studies methods with MATLAB and uses the MSRC-12dataset andDHA dataset. The experiments of this paper suggest two conclusions:1) Under theskeleton feature, HCRF is superior to other models;2) Based on the HCRF, thefeatures extracted from depth information is better than the features extracted from theRGB information.