The Research On Human Activity Recognition System Based On Wearable Devices

With the development and maturity of advanced technologies such as the internet of things(IoT),artificial intelligence(AI)and cloud computing(CC),Human activity recognition(HAR)has become a research hotspot in the field of ubiquitous computing and pattern recognition.Human activity recognition system includes four parts"perception","access","platform",and "application".There are two main ways of obtaining activity information in HAR:vision-based HAR and wearable-based HAR Studies have shown that vision-based activity recognition has certain shortcomings in terms of privacy and space-time applicability.In addition,the development of intelligent hardware also provides a good opportunity for activity recognition.Nowadays,human activity recognition technology based on wearable devices has been initially applied in some fields such as human motion analysis,smart home,human-computer interaction,medical diagnosis and monitoringAt present,human activity recognition system based on wearable devices still have certain shortcomings in terms of wireless access and platform algorithms.In terms of wireless access,wearable devices have extremely high power consumption requirements,at the same time,in order to achieve long-distance real-time recognition,how to achieve low power consumption and long-distance transmission is a great test Wireless access technologies mainly include Bluetooth and 4G,etc.,however,it is difficult for these technologies to achieve a balance in terms of power consumption and distance.In terms of platform algorithms,the mainstream algorithm models just consider whether human activity contains some features but ignore the positional relationship between these features.On the other hand,they have low accuracy of activity recognition for confusing activity.Therefore,it is great practical significance to make the system more perfect by effectively improving the system access and platform algorithms.Concerning the actual needs of human activity recognition systems based on wearable devices,this thesis studies data transmission and human activity recognition algorithm models.The specific research contents are as follows:(1)This thesis studies the transmission of human behavior recognition system based on wearable devices,and proposes a human activity recognition system based on long range(Lora)technology for the problem of compatibility of transmission distance and power consumption.(2)This thesis proposes a human activity recognition system based on capsule for the problem that the current framework cannot identify the spatial relationship between the features.The convolutional neural network(CNN)is encapsulated into capsules in parallel,the features are identified by each convolution kernel,and the vector relationship between features is represented using parallel convolution kernels.(3)The platform algorithm uses the same parameter set to process all kinds of input activities,resulting in lower accuracy of confusing activity recognition.This thesis proposes an adaptive multi-state pipeline framework based on set pair analysis(SPA),which is preprocessed by machine learning in the main pipeline,and then SPA technology is used for classification.Enter different sub-channels for classification processing through preset parameters,and finally perform fusion through the fusion pipeline.Through the above study,this thesis not only makes it possible to realize real-time human activity recognition based on the wearable device by solving the compatibility between the power consumption and distance,but also improves the overall performance of human activity recognition systems by optimizing platform algorithms.