A Gesture Recognition Method Based On CNN And OSELM

With the continuous deep research of artificial intelligence,human-computer interaction is getting hotter and hotter.Computer vision has gradually become the mainstream of current research,and more and more people have joined the field of research on human-computer interaction.Different researchers have studied different ways of interaction,such as face,gestures,body posture and so on.The main medium for studying human-computer interaction is to study different Convolutional Neural Networks(CNN)or to use different machine learning methods.In the research process of gesture recognition,the traditional method is to divide the gesture from the picture by segmenting the gesture,extract the Hu moment feature of the gesture and describe the contour of the finger,and put the extracted information into the information.BP neural network for training recognition.This method is slightly lacking in the accuracy of recognition.Combining convolutional neural network with online Sequential Extreme Learning Machine(OSELM)to achieve high recognition rate.To this end,the following work has been done:Firstly,The training data set is prepared for the subject to be studied.The research content is self-made gesture training set,and the verification experiment is performed on the JTD static gesture graph data set to improve the robustness of the research content.In the feature extraction of gestures,the VGG16-like network model in the convolutional neural network is used to train the gesture dataset while extracting gesture features.Secondly,The extracted gesture features are placed in the Online Sequential Extreme Learning Machine(OSELM)training recognition,and the different classification recognition algorithms are compared,and the different classifiers are compared in feature extraction.The advantages and disadvantages of the aspects and the different classification algorithms in the recognition of the correct rate.At the same time,the online sequential extreme learning machine has faster training speed and higher recognition rate than other algorithms on different size samples,which provides more abundant data forms for gesture recognition.Through the comparison of experimental data,this paper proposes that the method of combining neural network and online sequential extreme machine can be applied to gesture recognition with higher recognition accuracy than simple neural network,and it also has better applicability in different size data.Finally,a gesture recognition system is developed for the proposed algorithm.The algorithm is applied in real life to realize the landing application of the algorithm.Combining the algorithm with game animation,the practicability of the proposed algorithm is verified by practical application.