Recognition Of Human Continuous Action With 3D CNN

The recognition of human continuous action is the basis of human-computer interaction of service robot that plays a crucial role in the development of service robot.At present,a large number of researches on action recognition mainly focus on one single human action performed by one single person,and there are a few studies on the recognition of continuous action.Meanwhile,the development of action recognition is restricted that traditional machine learning is difficult to extract action sequence feature.In recent years,there has been a gradual rise of deep learning which can automatically learn the features of sample sets that arise a boom of research in all fields.In this paper,we combine the hottest current deep learning with the recognition of human continuous action that design an improved 3D CNN mixture model to provide a solution for the problem of human continuous action recognition.Aiming at the problem of difficult recognition of human continuous action,this paper carries out a series of preprocessing to the action sequence before the network structure design,and extracts the gray feature,motion feature and edge feature of the original sample respectively.In terms of motion feature extraction,L-K optical flow method which is improved by convolution kernel,because of the computational complexity of dense optical flow and L-K optical flow method,which is difficult to capture large-scale motion.In the aspect of edge feature extraction,since the resolution of the image needs to be unified before the network training,the edge feature ambiguity will be caused during the resampling process.Therefore,the edge feature channel is added that the Gabor filter is applied to prominent of the texture of the image sequences.After the image preprocessing forms a multi-channel,the 3D CNN network structure is designed for each channel.If the number of channels is too large,the network characteristics of each channel are directly connected which the amount of computation is too large and some features are blurred.Therefore,using the discrete wavelet transform to integrate the edge feature channel and the motion feature channel which improve the overall network performance.After extracting features through deep network,a number of classifiers are designed to identify and classify the action sequences.The mixture model combining 3D CNN and SVM is proved to be superior to action recognition.Also,we visualize the characteristics of each network layer to understanding of the operation of the network layer.In this paper,the mixture model designed is applied to the human continuous action recognition.Since the training samples are segmented actions,no pre-division action is needed in the training process and only a certain number of frames can be collected to identify which provides a good solution for human continuous action recognition and also have a good practical significance.