| In recent years,with the popularity of smart terminals and mobile communication networks,various types of video social entertainment software have risen sharply,accompanied by massive and complex video data.How to process these information data and use computers to analyze and understand large-scale video data intelligently is particularly important.In various types of videos,human actions often contain the main information of video.Action recognition and temporal action detection are also the most important research topics in video understanding.The focus of this thesis is on temporal action detection in video understanding.This thesis is mainly based on deep learning technology.From the perspective of the use of context information and the processing of video features with different spans,context cascade network is proposed to solve the problem of inaccurate positioning of motion boundaries caused by large spans and and blurred boundaries of video action instances.Specifically,there are the following points:(1)From the perspective of temporal context,this thesis continuously changes the length of the context and raises the threshold through a multi-level cascade to establish a context-sensitive multi-level temporal action proposal model to accurately and effectively locate the boundaries of the action.(2)From the perspective of video feature processing,this thesis uses feature pyramids combined with context cascade networks to align features and utilize features with different resolutions to locate action boundaries.(3)For the proposal classification part of the temporal action detection task,this thesis proposes a proposal classification model based on a long-time coding model in combination with a typical proposal classification model,which effectively improves the accuracy of action recognition.With all the above improvements,the model proposed in this thesis finally surpasses the current best method on the standard dataset of temporal action detection tasks. |
PDF Full Text Request