Research On Human Motion Recognition Technology Based On Kinect

Behavior recognition as an important technology in the field of human-computer interaction has attracted the attention of many researchers.The traditional behavior recognition is to extract the behavioral representation on the basis of color information for behavior recognition,but the behavioral representation based on color information is susceptible to factors such as illumination,background and human self-occlusion.Using the human skeletal joint data containing human body motion information to study the behavior recognition,to some extent,can avoid the influence of disturbing factors such as complex background,which brings new vitality to the study of human behavior recognition.At present,many research based on human joint information is to identify the behavior of a single action instance,but there are relatively few studies on the recognition of action sequences that contain multiple action instances.From the perspective of the data category being studied,most research is to extract the behavioral characters using color information,depth information and human joint information respectively.Then,the feature fusion is used to identify the characters,ignoring the supplementary information about the type of handheld objects that is used to identify the interactive behavior.Firstly,this thesis proposes a skeletal feature based on joint information,which is called the action coding map.Then the support vector machine(SVM)is used as the behavior characteristic classifier.The experiment on MSCR-12,which is a Microsoft’s public action data set,proves the validity of the behavior representation of the action encoding map.The influence of the characteristic parameters of the action code map on the behavioral characterization of the feature is analyzed.For a sequence of actions that contain multiple action instances,this paper establishes an action sequence segmentation model based on GMM algorithm.First,the action coding map for the whole sequence of actions is obtained.Then the corresponding action coding mask map is generated by using the threshold segmentation method on the basis of the action coding map.Next,on the basis of the mask map,the eight-direction-seed algorithm is used to extract time segments during which the human body postures are unchanged approximately,and the GMM algorithm is used to judge the time segments between the action instances.Then these stationary time segments are regarded as the segmentation accordance to separate the multiple action instances from the action sequences.Finally,the action coding features of each action instance are extracted,and the action recognition process of the single action instance is completed by the support vector m achine,thereby,the action recognition of the action sequence is realized.For human interaction,the type of handheld object has an important information supplement to the recognition of behavior.This paper first designs a solution to detect whether objects are held in the hand with advantage of the Kinect platform which can capture the color,depth,and joint information of objects at the approximately same time.If an object is held,the local color image of the object is obtained at the same time,and a complete color image database is established.The category of object is determined by using the Bag of Feature(BOF)image retrieval algorithm based on SIFT feature.Finally,the class of the object is fused with the ACM feature to be used to identify behaviors with support vector machine.Thereby,the purpose of behavior understanding is realized.