Research On Human Motion Recognition Algorithm Based On Neural Network

Nowadays,human motion recognition has become an important research topic in the field of computer vision.It has wide application and research values in human-computer interaction,virtual reality,behavior detection,video surveillance and many other directions.The most primitive method of human motion recognition mainly relies on workers sitting in the monitoring room all day to distinguish the human actions on the screen.This detection method is extremely resource intensive.And because of the difference in subjective consciousness,each person has his own unique judgment on the same event.Moreover,even the same person,the judgment of human actions will be affected by emotions and various conditions.The traditional motion recognition methods mostly pre-design features based on the corresponding data set,and then use the already set features to identify the data set.This method requires a lot of resources and is relatively unsatisfactory.Therefore,it is necessary to have an end-to-end learning algorithm that can automatically learn features.The deep learning algorithm in the field of machine vision just meets this point.It can not only tirelessly learn the action features in the automatic learning data,but also can be used to accomplish tasks similar to human eye observation and judgment.In this paper,deep learning is applied to human motion recognition tasks,which can effectively improve the recognition accuracy of human motion,and finally complete the human motion recognition task in the video.In order to overcome the shortcomings of original artificial recognition and traditional algorithms,the main research contents and innovations of this paper are as follows:(1)The traditional way to process video data is to extract the RGB frames of the data and then input them into the algorithm for training.This method of training only uses RGB images focuses on the spatial position structure of human motion in video but ignores human motion is a continuous motion,missing the continuity of motion.Therefore,the processing of the input data needs to maintain time continuity.In the second chapter of this paper,the video data is processed,not only the RGB form of the data is extracted,but also the optical flow form is extracted.This method preserves the temporal continuity of the data for better recognition.(2)The video human motion recognition technology based on the traditional method needs to design the corresponding feature template for the fixed action,which is timeconsuming and labor-intensive,and the classification effect is not good.In the fourth chapter,the paper studies the human motion recognition algorithm based on batch normalization twostream neural network.The network structure combining batch normalization convolutional neural network and single-layer Long Short-Term Memory neural network(LSTM)cyclic neural network.This method is used to automatically learn the underlying characteristics of the data.Experiments show that the algorithm achieves better accuracy than traditional methods.(3)In the fifth chapter of this paper,a model combining convolutional neural network and two-layer LSTM cyclic neural network,which combines shallow features and high-level features,is proposed.The model also uses the optical flow image and RGB image extracted from the data set as input,and the final recognition result is obtained by weighted fusion of the result model.Experiments show that this algorithm also achieves better accuracy than other excellent algorithms.