Research On Video Action Recognition Based On Improved Predictive Recurrent Neural Network

With the continuous development of information technology,people pay more and more attention to action recognition technology.Action recognition technology plays an important role in gesture recognition,human-machine interaction,abnormal behavior detection,intelligent video surveillance and other fields.Therefore,the research on video action recognition has strong practical significance.At present,most of the algorithm research on action recognition is based on deep learning method,and most of them use supervised learning for training,which relies on large-scale labeled data sets,and the method of obtaining data features through supervised learning does not conform to the reasoning that humans have predictive ability.Through unsupervised learning network,we can better obtain the potential features in the data and optimize the performance of the network.This paper takes the behaviors in the video as the research object,improves the predictive recurrent neural network(Pred RNN)action prediction model,builds an action recognition network,and designs a video action recognition system on this basis.The main research contents of this paper are as follows:(1)Research on video prediction algorithm based on adaptive feature fusion.Aiming at the problem that the gradient of the spatiotemporal memory flow structure of Pred RNN network disappears with the increase of network depth,resulting in the fuzzy result of long-term video prediction,a video prediction algorithm based on adaptive feature fusion is proposed to fuse the features of spatiotemporal memory information,increase the gradient of deep network,improve the accuracy of video prediction algorithm,and solve the problem of fuzzy result of long-term video prediction.(2)Research on action recognition algorithm based on video prediction.Aiming at the problem that the construction of action recognition network using supervised learning is limited to large-scale annotation data sets,an action recognition model based on video prediction model is proposed.Using the action prediction model based on unsupervised learning training to obtain the key characteristics of action,get rid of the limitations of supervised learning,and solve the problem of the dependence of action recognition network on large-scale labeled data sets.The feature information is fused from the time dimension to better reflect the relationship of feature information in the time dimension,enhance the robustness of action recognition network to the speed of action,and improve the performance of video action recognition network.(3)Design of video action recognition system.Based on the above related methods and technologies,the video action recognition system is designed,including front-end interaction module,data preprocessing module,action recognition module,feature visualization module and data management module.Split the input video into image frame sequences,and perform feature extraction and classification on the image frame sequences.Through the experimental test of the action recognition system,the validity of the action recognition algorithm and action recognition system constructed based on the video prediction model is verified.