Research On Human Action Prediction Based On Vision

With the advent of artificial intelligence and 5G,human action prediction has attracted more and more attention.Compared with human action recognition,human action prediction has the characteristics of timeliness and effectiveness and has broad application prospects in the fields of automatic navigation,video surveillance and early warning,and robot human-computer interaction.However,the current research on human action prediction is hampered by the lack of observational information and confusion of actions,which brings great challenges to the development of human action prediction.Given the above problems,we utilize the methods of deep learning to study the related technologies of human action prediction from the design of the early action prediction network model,the design of the early action prediction loss function,and the dense action anticipation method.The main research contents are as follows.First,design of early action prediction network model.Aiming at the problem of lack of observational information in the task of early action prediction,a novel method,which is termed feature map-based hierarchical attended spatial temporal graph convolutional networks,is constructed by taking the hierarchical spatial temporal graph convolutional network structure as the main body,supplemented by the attention module based on feature maps.Experiments are performed on NTU RGB-D 60 dataset and SYSU 3DHOI dataset.Second,research on early action prediction based on loss function.Aiming at the problem of action confusion in the task of early action prediction,a loss function,which is termed reliability-based step loss function,is designed to guide the model to make accurate predictions as early as possible under the premise of sufficient confidence,and to distinguish confusing actions in the early stage of the action.Experiments are conducted on NTU RGB-D 60 dataset.Finally,dense action anticipation method.As an exploratory research goal in the task of dense action anticipation,a novel model,which is termed two-stream retentive long short-term memory networks,is designed to make use of high-level and low-level features at the same time,supplemented by a retentive memory module,to analyze the intricate relationship between actions,and finally make reasonable,long-term forecasts.Experiments are conducted on Breakfast dataset and 50 Salads dataset.