Research On Human Action Recognition Based On Hidden Markov Model

Pattern recognition has more and more applications in modern society,and human gesture recognition has always been research hot-spot in recent days.It's to automatically analyze videos using computer,for the purpose of making computer to have a human-like sense of the world.The technology can be used to detect any abnormal events in monitored places,or to analyze the action of monitored object.That's why it's so important.The paper will discuss human gesture recognition from following four steps: camera data collection and pre-process,fore-view detection,human feature parameter extraction and HMM human gesture recognition.Human gesture recognition discussed in this paper,is using chess-board method to locate the cameras,in order to get their distortion parameter and internal parameters.We use Mixed-Gaussian Model to create background model based on history video variations,and extract fore-view information from it in order to do background reduction.We use light-flow method to do fore-view detection in order to get motion feature and light-flow info of target person,and we compared the effectivity of sparse and dense light flow.We do the feature extraction of human gesture from two aspects: normal geometrical feature and light flow information of human.On geometrical feature,we extract the width-height ratio,perimeter-size ratio,mass point,exzentritaet eccentricity and feature angle.On light flow information,we created the descriptor of light flow using mesh-based method.We also do feature merge of the above two,in order to create word-bag model.The paper discussed using HMM model to recognize human gesture,and train the HMM parameters on four different ordinary gestures.Detection of all four gestures are implemented.We use the videos shot indoor at night as library to do the experiment.Result shows high correctness rate and good anti-noise feature.