Action Recognition Based On HOG/HOF And Space-time Interest Point

The behavior recognition of the video surveillance system identified as an important intelligence applications, has now become a very important field of computer vision research direction. Behavior recognition process can be divided into the following four parts:target detection, target tracking, target classification and target behavior recognition.This thesis is based on pattern recognition and digital image processing theory, through analysis and research on domestic and international recognition method on human behavior, proposed Histograms of Oriented Gradients (HOG), Histograms of Optical Flow (HOF) two-dimensional global features combined with motion feature based behavior recognition methods, research and implemented HOG3D based dimensional local time-space interest points feature description of behavior identification method. HOG, HOF feature based behavior recognition method firstly using frame difference method and the threshold method to detect the motion region, extracted human motion region, Secondly described the extracted motion area with HOG and HOF features, and extract the center of gravity of a binary image which is used to detect motion area, through the center of gravity movement judgment obtained velocity characteristics. Thirdly, using K nearest neighbor algorithm test HOG and HOF feature of all data with sample data to obtain the category of each picture tested. Finally, make all data of each image as an representation of category number, take20category number as an serial, using Hidden Markov Model to classify all serials, take the result of HMM as the final result. HOG3D feature based behavior recognition firstly packet the video frame then use the Harris-Corner3D detection algorithm to detect3D interest point in video sequence. Secondly, removing the interesting point which is not meet the condition, described the rest point using HOG3D feature describer. Thirdly, using K-means clustering algorithm to cluster all interest point, the cluster center can be defined as the dictionary arsenals of bag of words model. Finally, statistical the interesting point which is nearest the center of dictionary of each video sequence, describe the sequence with a center histogram. Test all the video sequence histogram with support vector machine using cross-validation method.This thesis is based on KTH dataset selected25people in four different scenarios doing six actions. Research and simulate the human action recognition algorithm in this thesis. The final result proved that this human action recognition algorithm is useful and effective.