Analysis And Research Of Temporal Event

With the development of global technology and the maturity of deep learning technology,the Internet has been in the era of artificial intelligence.Computer vision is an important field of artificial intelligence research.The analysis and research of temporal event has become a hot spot at present,and has broad market and application prospects in many fields such as automatic driving,intelligent security,short video recommendation,and action behavior analysis.At present,temporal event analysis and research has entered a new stage.From the traditional methods in the past to the popular neural network,although the performance on each dataset are constantly improving,there is still a certain gap from the actual scene application.The main difficulty lies in the complexity and diversity of the temporal event task.It requires the identification of the action categories in temporal event,the detection of the action range and the determination of the action duration.Moreover,due to the complexity of the background,the magnitude of the freedom movement,the difference in the duration of the event,and the low resolution and ambiguity caused by the jitter of the camera when shooting video,the great difficulty is also brought to the analysis and research of temporal event.For the complexity and diversity of temporal event tasks,this work uses hybrid convolution and attention mechanisms to identify actions in temporal event,combine P3 D and Faster RCNN to detect targets in temporal event,and improve boundary-sensitive networks to determine action durations and action proposal in the event.The main three parts of this work as follows:1.A method for action recognition in temporal event based on hybrid convolution and attention mechanism is proposed.The algorithm takes the whole short video in the temporal event as the processing object.Firstly,the short video is divided into multiple random samples,then the P3 D convolution feature is extracted from the video segment,and the extracted video segment features are combined to obtain the mixed convolution feature map.The attention-mechanism operation is performed on the mixed convolution feature map,and the attention mechanism map is obtained,which is transformed into an attention descriptor,and then the whole video feature is characterized and used for action recognition in temporal event.2.A method based on P3 D and Faster RCNN for action detection in temporal event is proposed.Based on the Faster RCNN object detection framework and the characteristics of continuous change in action,this algorithm P3 D convolution is used to analyze the human action feature in key frame and effectively combined in the RPN network of key frames.When the small changes of the position in action are ignored,the understanding of the action detection by the Faster RCNN is enhanced,and the accuracy of the action detection in the temporal event is improved.3.A method for action proposal in temporal event based on bidirectional LSTM to improve boundary-sensitive networks is proposed.This algorithm effectively improves the method of action proposal in boundary-sensitive network.After extracting the two-stream video feature,it is divided into three parts in the temporal evaluation module,and the 1D timing convolution is used for the start frame and the end frame,bidirectional LSTM is used for the intermediate frame,and then the probability of the start and end frames is normalized again in the propsal generation module.Finally boundary-sensitive feature is constituted,and more accurate temporal event action proposal is obtained after non-maximum suppression.