Research On First-view Video Action Recognition Technology Based On Deep Learning

The head-mounted camera has been widely used with the development of portable technology.It provides a large number of video data from the first perspective.Human action recognition in the first person video is paid more attention.Human action recognition technology has important applications in virtual reality,intelligent video surveillance,smart home,video security.It has great demand in the field of human-machine interaction in robots technology and so on.In the fields of image recognition,detection,and so on,the method of deep learning is widely used,and the performance is improved obviously.Based on the deep learning framework,in order to solve the video action recognition from the first perspective,we propose an improved algorithm of deep multi network fusion for human-machine interaction video action recognition and outdoor action recognition.First of all,we propose a structure of global and local 3D convolution network fusion,the global 3D network convolution focuses on capturing the observers' motion while local saliency 3D convolutional network can capture the actors' action.Secondly,we propose a fusion structure of scene 2D convolution network and global 3D convolution network,the global 3D convolution network assisted scene convonlution network can achieve good representation of the outdoor action and improve the accuracy of action recognition.Finally,we design and implement a system based on deep learning for first person action recognition,it consists modules of the user selection of video as input,video preprocess and algorithm model selection,the final fusion results,convenient for the user to analyze the results of the practical application of the algorithm.