| Action recognition is a hot research direction,which has been widely used in various intelligent fields based on video content.Among them,individual action recognition has received extensive attention and research due to its practical value.In the field of public safety,it is of great significance to accurately and quickly analyze the behavior of individuals in videos for finding abnormal individuals in pedestrians and maintaining social order.In the current research on individual action recognition,the SlowFast model has become one of the most popular recognition algorithm models by its two-stream structure that integrates spatio-temporal information and good end-to-end characteristics.Since the model is modeled by Convolutional Neural Networks(CNN),the overall modeling ability of the model for long time series video is insufficient,which limits the application scope of the model.In addition,the SlowFast model focuses on local features,and the extraction of global interaction features is incomplete,which leads to the low classification accuracy of the model.Therefore,in view of the above two problems,this thesis carries out the following work:1.A SlowFast algorithm model fusing time series segmentation,called TSSFN model,is designed.The model maintains the overall structure of the original SlowFast model,and integrates the idea of time series segmentation.Through sparse sampling and time domain segmentation,the feature information of the whole video is extracted and fused to realize video-level modeling.Finally,the experimental results show that the proposed algorithm model has good modeling ability for long time series videos compared with the original SlowFast model,which expands the application scope of the SlowFast model.Compared with other models,the accuracy of individual behavior recognition is improved,which verifies the success of the model.2.An individual action recognition model integrated with self-attention mechanism,namely AF-TSSFN model,is designed.In this thesis,the Self-Attention mechanism is integrated into the TSSFN model to extract global features and contextual interaction features.And Fully Convolutional Networks(FCN)is used to improve the classification performance of the model,so as to achieve the purpose of accurately identifying and classifying individual behaviors.Finally,ablation experiments and comparison experiments show that compared with the TSSFN model,the AF-TSSFN model proposed in this thesis can more effectively combine the context information,extract the interactive features and global features of the actions in the video,and more accurately complete the individual action recognition in the video.And the accuracy of individual behavior recognition of the model is improved compared with other mainstream models,which verifies the superiority of the model.The individual behavior recognition model proposed in this thesis is an improvement on the existing network classification model,which is an expansion of the application scope of the SlowFast model.It has certain academic and application value in the field of video understanding,and provides a good reference value and academic foundation for accurately predicting and intervening harmful behaviors and maintaining social security. |
