Research On Human Behavior Recognition Based On Improved CNN-LSTM

In recent years,human behavior recognition has been widely used in the fields of computer vision and human-computer interaction.With the rapid expansion of the smartphone market and the rapid development of micro-sensors,a variety of MEMS sensor devices are embedded in smartphones.The research on human behavior recognition based on sensor-based intelligent devices has become a new branch,which provides a good support in health care and medical detection.Although this research has achieved good results in recent years,there are still many deficiencies,such as the privacy of personal data is not easy to access.Because of poor data quality collected by individuals,the precision of motion recognition rate is still not sufficient.In the domain of behavior recognition algorithm investigation,the feature extraction project in machine learning requires the researchers to have some prior knowledge and a lot of human interaction.With the advent of neural network,human behavior recognition algorithm has been transformed from traditional machine learning to neural network,which automatically selects features through training networks to achieve classification.This thesis focuses on human behavior recognition based on neural networks and smart devices.The test datasets come from three public datasets(UniMiB SHAR public datasets,MobiFall public datasets,MobiAct public datasets)covering falls and daily behaviors.Through analyzing and processing the behave data sets,which are combined with the current popular neural network for classification and modeling.The contribution of this thesis including three parts,which are the pre-processing of sensor-based human activity data,the recognition and classification of human behavior activities,algorithm evaluation and system implementation.In the first part,the public datasets need to data analysis and noise reduction processing,after data analysis,we use the moving average filtering to reduce noise.In order to facilitate the training and classification of neural networks,this thesis also compiled a set of data format template processing scheme.In the second part,a new neural network model is designed to classify human behaviors according to the characteristics of data.Among them,the bidirectional long short time memory network can effectively extract the hidden time series features.In this part,the improved convolutional neural network and the bidirectional long shortterm memory network are designed for multi-feature fusion,and the algorithm accuracy is gradually improved through continuous training of the network.In the third part,the proposed neural network is compared with other algorithms for human behavior recognition(including machine learning algorithms and neural network algorithms).Experimental results show that the method can improve the quality of human behavior recognition and has higher precision and good universality in a wide variety of scenarios.Finally,the human behavior recognition system has been implemented in the practical application.