Research On Human Behavior Detection Method For Mine Videos

In recent years,human behavior detection becomes a hot area of study in computer version.It analyzes the video data with image processing,pattern recognition,so as to build the relationship between the raw data and high-level semantics.It is widely used in biometric identification technology,human-computer interaction,content-based video retrieval and so on.Human behavior detection can be applied to mine videos to detect abnormal behaviors of miners.Based on results of the detection real-time alarm and related operation can be realized,and the mine accidents can be reduced.The thesis studies the human behavior detection in mine scene and hopes to classify simple human behavior from some mine video.Based on the current research findings on detection of human behavior,the thesis divides the detection of human behavior into extraction of body motion area,extraction and express of human behavior feature in a video,classification of human behavior.(1)The thesis proposes a body motion area extraction fusion algorithm which merges the original Adaboost algorithm based on Haar features and temporal difference method.Firstly,the thesis uses temporal difference method for background modeling to obtain initial body motion area,and then ueses the classifier which is generated by Adaboost algorithm to get more accurate body motion area.In this way,we use temporal difference method for background modeling,and then use the classification which is generated by means of Adaboost algorithm to get more accurate body motion area.In this way,the influences from complex background can be reduced.(2)The thesis uses Harris3 D Corner detection algorithm to detect STIP in mine videos,calculates HOG and HOF features in the neighborhood of STIPs,and then normalizeds all the HOG and HOF features of each STIP.(3)The thesis builds bag of feature model and classify different behavior using Support Vector Machine(SVM).In order to get the dictionary of bag of feature,the thesis uses K-means algorithm to cluster all the HOG/HOF features of training set.By calculating Euclidean Distance between HOG/HOF features of a video and the dictionary of bag of feature,visual vocabulary histogram which will be converted to feature vector can be generated.All the feature vectors will be used to train SVM classifierwhich can be used to detect human behavior.Resource Description Framework(RDF)is used to express and store human behavior classification.