| With the rapid development of computer science and Internet,computer vision has received an extensive attention and research in various fields.Especially the action recognition based on the video plays a significant role on the applications of surveillance and security.Due to the rise of deep learning,the performance of computer vision algorithms has made a breakthrough.In the area of action recognition,RGB-D camera provides the depth information of action videos and also brings the challenges of information processing.At the same time,video feature mining remains one of the difficulties in action recognition.In view of these problems and difficulties,this paper has done the following work:We first propose an approach for multi-modal action recognition based on the deep neural networks.In order to processing different modal video information,different artificial networks are utilized and combined to exploit the multi-modal features.Secondly,to address the lacking of saliency due to complex scene information of video,we propose an action recognition method based on visual saliency.This approach combines the 3DCNN and RNN as an end-to-end model to keep the temporal and spatial sequence features of actions.At last,We design and implement the Web-based action recognition system.Users can choice videos on the website through the system.After the analysis and recognition of the action recognition models on the server,the system can predict the classify results on the user-side,which will help them to analyze the actual results. |
