Study On The Action Recognition Based On Multi-layer Recurrent Neural Network

With the rapid changes in social life,intelligent monitoring system is gradually becoming the focus of attention.Intelligent monitoring system refers to the real-time monitoring of people,things and objects in the specific scenes.At the same time,the system gives the analysis and prediction of specific events that may occur in these scenes,especially the automatic recognition of human actions to achieve the tracking of specific people or early warning.Human action recognition based on video is an important part of the intelligent monitoring system,and it is also an active field of computer vision.The intelligent monitoring system which applies human action recognition can not only filter low-risk information in the scene during safety monitoring,improve the efficiency of safety monitoring and lower labor costs,but also play an important role in smart homes and smart old-age care.By using this technique,it can give health risk management and timely warning of sudden illness which can effectively solve the disability or death caused by not being treated in time.This paper mainly explores human action recognition based on the video images.We adopt the structure of the two-stream network,and improve the way to handle the input data.Our model has applied the object detection network which only focus on the human actions in the foreground objects.in addition,a novel multi-layer recurrent neural network is proposed to recognize human actions.This paper uses the NTU RGB + D human action recognition dataset,and we chose seven classes in it,such as brush hair,sit down,stand up,hand waving,falling down,headache and neck pain.The main contributions are as follows:(1)Base on the traditional two-stream,we introduce an object detection network to two-stream,and propose a model of human action recognition based on multi-layer recurrent neural network.Our model uses pyramid three-dimensional dilated convolution to process continuous video images,and combined with Convolutional Long Short-Term Memory(Conv LSTM),which provides a pyramid convolutional long short-term memory network that can analyze human actions in real time.(2)To solve the problem that the current two-stream cannot read continuous video image sequences as input,we use three-dimensional convolution to improve it.Threedimensional convolution has excellent ability to extract spatial information in video sequences.At the same time,three-dimensional convolution has the ability to read continuous video image sequences.By using three-dimensional convolution,it can simply and effectively solve the problem of reading continuous video images.