Research On Video Action Recognition Based On Improved Long Short-term Memory Network

Human action recognition always is a research hotspot in the field of intelligent analysis about videos.In recent years,it has been widely used in many aspects,like virtual reality,human-computer interaction and short video shooting,and has great research significance.With the rapid development of deep learning algorithms in computer vision,action recognition algorithms based on deep learning emerge one after another.Researchers innovate action recognition algorithms from the aspects of recognition accuracy,recognition speed,model structure and so on.However,the current action recognition algorithms still have room for improvement in recognition accuracy and model complexity.Thanks to the rapid development of deep learning,especially Long Short-Term Memory network in the field of action recognition.In this paper,on the basis of Long Short-Term Memory network,the existing action recognition networks are improved to achieve further research on action recognition algorithms.The main work of this paper is as follows:(1)aiming at the problem that the traditional Long Short-Term Memory network only pays attention to the characteristics of time series,on the basis of the traditional Long Short-Term Memory unit,this paper introduces the concept of differential control in PID control into the deep learning network and propose Long Short-Term Memory units with input differential.This method can not only increase the influence of short-term time series on action recognition,but also consider the influence of different velocity and acceleration of human body on action recognition,in which the first-order differential corresponds to the action velocity and the second-order differential corresponds to the action acceleration.In this paper,the input differential Long Short-Term Memory units are used to replace the basic Long Short-Term Memory units in the Long Recurrent Convolutional Network,and verification experiments are carried out on three commonly used action recognition data sets.(2)by studying the influence of time-space differential on action position,and considering the influence of action speed and position change on action recognition,this paper analyzes the spatio-temporal Long Short-Term Memory network from the point of view of control.it is considered that it can be added links to strengthen differential features,so as to enhance the characteristics of movement speed changes and position changes.First of all,on the basis of spatio-temporal Long Short-Term Memory memory network,this paper proposes an improved spatio-temporal differential l Long Short-Term Memory unit,which can take into account both spatio-temporal information and limb position change information with time.Secondly,this paper further designs a stacking method suitable for the spatio-temporal differential Long Short-Term Memory unit,that is,it increases the horizontal transmission of spatial memory state on the basis of the vertical transmission of spatio-temporal Long Short-Term Memory units.In this paper,the stacked spatio-temporal differential Long Short-Term Memory units are tested for accuracy and stability on different data sets,and compared with other action recognition algorithms based on deep learning.