Research On Human Action Recognition Base On Deep Learning

A video-based human action recognition is widely used in various fields such as intelligent monitoring,security,content-based video annotation and intelligent human-computer interaction.It can solve the huge human resource consumption problem caused by traditional manual-based video processing.And at the same time,automated human action recognition also solves input problems in smart human-computer interaction.Traditional human action recognition is based on the image processing and computer vision technology,which is suffering the disadvantages of low accuracy,a large amount of calculation,complicated algorithm logic and high background knowledge requirements for algorithm researchers and engineer.In recent years,the deep learning algorithm has made a breakthrough in various fields of computer vision.There are many advantages in introducing deep learning algorithms into the field of action recognition.The deep learning based feature extraction algorithm can overcome some of the deficiencies of traditional artificially designed feature extraction algorithm.It improves the accuracy of the action recognition and greatly simplifies the algorithm design as well as the implementation process simultaneously.In order to study the performance of deep learning algorithm in action recognition further,this paper takes the optimization of deep learning network structure as a starting point,designing and implementing two kinds of action recognition algorithms based on deep learning.These two algorithms are based on the 3D convolutional neural network as well as the Long short-term memory network implemented for sequence processing.The algorithm of 3D convolutional based network uses multiple small convolution kernel stacks instead of large convolution kernels,which deepening the neural network's depth while reducing the network parameter amount.And then,the state-of-the-art batch normalization algorithm is used to improves training speed.The long-short-term memory network based action recognition algorithm uses the 3-dimensional convolutional network implemented in this paper as a feature extractor of pre-existing spatial domain and time domain.Then we expand the output feature map of the 3-dimensional convolutional network into feature vectors and use the long short-term memory network to extract the frame-related information contained in the video sequence.Finally,we verify the performance of the algorithm in the UCF-101 dataset and HMDB-51 dataset.The highest accuracy of our algorithm on UCF-101 dataset can reach up to 94% and on the HMDB-51 dataset it can reach up to 68%.This research work proves the feasibility of 3D convolutional network in the application of behavior identification.At the same time,it demonstrates the superiority of the fusion algorithm based on different deep learning networks for video sequence processing.