Research On Human Motion Recognition Based On Image

Human motion recognition is the key research direction in the field of computer vision at present,which has practical value and broad application prospects in many fields such as security monitoring,intelligent medical care,virtual reality and so on.Because the traditional human motion recognition method needs to select features manually and choose a suitable classifier,the classification effect is not good and lacks real-time,so this thesis mainly uses convolution neural network to study human motion recognition method.Convolutional neural network no longer needs to manually select features,but automatically extracts image features from low level to high level through end-to-end training,which improves the recognition efficiency and accuracy.In this thesis,UCF101 data set is used to study a variety of human motion recognition algorithms based on convolutional neural network.The main research contents include the following aspects:(1)Firstly,the method of human motion recognition based on convolutional neural network is studied.Because 3D convolutional neural network can directly extract the features of video time series,the method of human motion recognition based on 3D convolutional neural network was first studied,but 3D convolutional neural network can not directly use the method of transfer learning,so 2D convolutional neural network was studied.In order to add time information to 2D convolutional neural network,this thesis adds LSTM long-term and short-term memory network to 2D convolutional neural network,which makes image sequences have spatio-temporal correlation and solves the problems of gradient disappearance and gradient explosion in the process of long sequence training.On this basis,the ResNet-152 residual network structure model pre-trained from ImageNet data set is added to improve the recognition accuracy.(2)Secondly,the method of human motion recognition based on two-stream convolutional neural network is studied.The model pre-trained from ImageNet data set is used as the network model,the video single frame of UCF101 data set is sent to spatial stream to extract image environment and human spatial features,and then stacked optical stream images are sent to temporal stream to extract dynamic temporal features of human motion.The experimental results show that two-stream convolutional neural network can indeed improve the generalization ability and recognition effect of the model.(3)Finally,the method of human motion recognition based on two-stream inflated 3D convolution neural network is studied.The two-stream inflated 3D convolutional neural network combines the advantages of 3D convolutional neural network,two-stream network and transfer learning,and adopts GoogleNet network structure.The network model used is pre-trained from ImageNet and Kinetics data set and then transfered to UCF101 data set,and at the same time,it also uses video single frame images and stacked optical flow images for two-stream training,which improves the recognition effect obviously.By comparing the above five algorithms in UCF101 data set,the recognition algorithm based on two-stream inflated 3D convolutional neural network has the best effect,and the final average accuracy rate reaches 97.09%.