Research On Human Action Recognition Based On Semi-supervised Co-training And Ensemble Learning

With the development of science and technology,human action recognition has been becoming an important research topic in the field of artificial intelligence and machine vision.It has been already applied in a lot of areas such as video monitoring,intelligent medical,sports analysis,man-machine intelligence etc.Meanwhile,because of the complexity of video,motion diversity,and a large number of labeled data are needed to train the strong recognition model,it faces the huge challenge.This thesis studies several topics in human action recognition,in particular,an unsupervised learning approach for analysis of human action recognition.Firstly,this thesis elaborates the goal and significance of background and research in human action recognition.Secondly,some key technologies in human action recognition such as key frame extraction,feature extraction,feature extraction and action recognition are briefly described.Finally,based on the current theoretical research on human action recognition,this thesis presents a human action recognition algorithm based on hybrid collaborative training and a human action recognition algorithm based on semi-supervised collaborative training and ensemble learning.A human action recognition prototype system is also developed.The main contents are as follow:1)We propose a human action recognition algorithm based on hybrid collaborative training.Aiming at the problem of insufficient of labeled data on the current human action recognition methods,we proposed a human action recognition algorithm based on hybrid collaborative training.Different types of recognition methods for action recognition field are employed in this method to build the base classifiers,which are then iteratively retrained to increase their generalization abilities.In general,our method can decrease the labeling cost and achieve complementary advantages of different recognition algorithms,improve the recognition performance.The experimental results show that the proposed algorithms can identify human action in the video more effectively.2)We propose a human action recognition algorithm based on semi-supervised collaborative training.With the increasing of iterations on hybrid collaborative training,The differences of base classifiers will get smaller and smaller.And base classifiers are underutilized.Aim at these problems,we proposed a human action recognition algorithm based on hybrid collaborative training and ensemble learning.This method set up a collection for each classifier,and add middle classifiers that are generated in the process of iterative training to their own collection.Then use this collection to select pseudo labeled data.In addition,we define a most evidence edge function which is based on confidence to select the pseudo labeled data.In the end,we use this method to identify human action.This method can overcome the problem that difference will get smaller and smaller in the process of iterative training,and further improve accuracy on human action recognition.3)A prototype system of human action recognition based on semi-supervised method is designed and implemented.We use C# that it is based on object-oriented language and MATLAB to program.By operating prototype system,result indicates that it can be used in human action recognition.Moreover,the prototype system functions well with a friendly interface and good maintainability.