Specific Action Detection Algorithm Based On Deep Learning

Human action analysis is one of the focus of exploration in computer vision,and its model and algorithm updates are changing with the progress of the Internet.Driven by the rich video resources,the accuracy of action recognition in the edited videos has been significantly improved.However,video data in real life is rich in content and often includes multiple types of actions,which are generally continuous and not edited.How to reasonably and effectively process the amount of video on the Internet is a very important research topic.Based on this,a very challenging task—temporal action detection,needs to detect the action segment in the unedited video data,that is,the proposal of the temporal action.Locate the start and end time of each action in the video.However,processing video data that contains multiple types of actions has certain complexity.There may be multiple action fragments and action types on the same video,so the study of temporal action detection is somewhat challenging.In order to improve the performance of temporal action proposal and detection,thesis proposed three efficient algorithms for temporal action proposal and detection.The main work is as follows:(1)Based on the efficient network structure of BSN(Boundary sensitive network)algorithm,a proposal generation algorithm network(Action Keyframe Connection Network)based on action key frame connection network,referred to as AKCN for short,is proposed.AKCN has constructed a new type of temporary action proposal generation network that combines Long Short-Term Memory(LSTM)and Convolutional Neural Networks(CNN)to model temporal and Effectively detects the start frame and end frame of the action.Enhanced the use of different temporal features in the entire network.The feature vectors extracted by this connection method are more discriminative and robust,which helps the network improve the accuracy and efficiency of the algorithm.(2)The attention mechanism is introduced on the basis of AKCN,and action key frames that introduce attention mechanism for specific action detection are proposed to connect to the network.The main function of the attention mechanism is to learn the similarity relationship between the various features,so as to give different features corresponding to the importance of the weight,that is,the important features will be given a larger numerical weight,for the current features that are not important Give a smaller weight.In this way,the detection accuracy can be significantly enhanced.In this thesis,the feature relationship module including attention mechanism is embedded in feature extraction and CNN network to remove redundant key frames.Experiments prove that the introduction of attention mechanism module makes the algorithm reach a higher level of efficiency and accuracy.(3)Based on the efficient AKCN algorithm,by further introducing the category judgment network,a specific temporal action detection algorithm based on the joint network is proposed.In the proposal generation phase,when the probability that the start frame and the end frame are merged into the same action is greater than a certain threshold,the author is allowed to generate an action proposal.The action proposal is used as the input of the existing action recognition network,and the video data is tested through the above algorithm to complete the task of building a specific action detection network.