Research Of Human Action Recognition Based On Voting Strategy With Multi-feature Classification Results

Human action recognition involves many disciplines,such as image processing,pattern recognition and artificial intelligence.It has been widely applied in many fields,such as family entertainment,video surveillance,virtual reality,human-computer interaction,and so on.At present,human action recognition is still at the stage of research and development.Although many human action recognition methods have been proposed,there are still four main challenges in human action recognition:(1)an action has not been clearly defined so that it can be distinguished from other actions;(2)the action data contains noise and redundancy,which will affect the accuracy and speed of action recognition;(3)a feature can describe a limited action categories,and there are no features that can describe all the actions;(4)there are major difficulties in recognizing interactive actions or group activities.In recent years,it's easier to extract 3D depth data from scenes with the development of the depth sensors.The 3D depth data contains more information than the traditional RGB image data,which helps to explore new action recognition methods.Therefore,the research of human action recognition based on the depth information is beginning to get hot.This thesis is based on depth data to explore methods to solve the above challenges(2)and challenges(3).Aiming at the problem of noise and redundancy in action data,a key frame extraction algorithm based on self-organizing mapping(SOM)neural network is proposed in this thesis.The competitive learning characteristics of SOM neural network are used to automatically cluster the data,and then the most representative data is selected from each class as the key frame data.Although the size of the keyframe data is small,it can accurately represent a series of gestures that make up the action.Using keyframe data for action recognition can reduce the impact of noise and redundancy on recognition results and recognition speed.For a single feature is only applicable to describe part of actions,this thesis proposes to describe human actions by combining multiple features.The method uses multiple features to describe the action from different aspects and makes full use of the complementarity between different features,making it more comprehensive to describe human actions.In order to further improve the accuracy of action recognition,this thesis puts forward a method of human action recognition based on voting strategy of multi-feature classification results in the stage of action recognition.It votes according to the classification results of different features,the action that obtains the maximum number of votes is the result of the final action recognition.For the case of equal number of possible votes,the action with the highest credit rating in the voted action categories is taken as the final action recognition result.By testing the key frame extraction algorithm based on SOM on the MSR DailyActivity3 D dataset,the results show that the proposed key frame extraction method can effectively reduce the adverse effects of noise and redundant data on the accuracy,and significantly improve the speed of human action recognition.Through the experiments of multi-feature voting strategy on the MSR Action3 D dataset and UTKinect Action dataset,the results show that the combination of multiple features to describe human actions is better than the single feature.The voting strategy of multi-feature classification results can improve the accuracy of human action recognition.Compared with other methods,the method of action recognition based on voting strategy of multi-feature results has a good recognition effect.