Research On Technologies Of High-accuracy And Low-cost Motion Mode Recognition In Wearable Computing

With the development of science and technology,China will usher in an informatization and intelligent industrial innovation in the future.This innovation is mainly based on the internet-of-things(Io T)and artificial intelligence(AI).Meanwhile,the popularity of Io T will also promote the development of edge computing.Since edge computing can effectively shorten the system delay,energy consumption,and protect the privacy of user data,it will become the main application field of the combination of Io T and AI.In this context of the development of edge computing,human motion mode recognition(MMR)based on wearable devices has become a hot research field in recent years and plays an increasingly important role in the fields of indoor positioning,intelligent software design,and elderly care.Even though many research results have been achieved in the field of MMR based on wearable devices,there are still a lot of problems that need to be solved.These problems include:(1)Incomplete datasets.The datasets used in previous work usually only contain a specific scene or consider a certain device usage,which are impossible to support robust MMR systems and unbiased evaluation.(2)There is a lack of systematic analysis of the key parameters in MMR systems.These parameters directly affect the accuracy and energy consumption of the entire system,but they have not been paid enough attention in previous work.(3)There is a lack of research on high-accuracy and low-cost MMR methods.Previous studies in MMR mainly focused on how to improve the recognition accuracy but ignored the size and energy consumption of methods proposed.However,high-accuracy and low-cost methods are precisely the key in realizing usable MMR systems.To resolve these problems,this paper conducts indepth research on high-accuracy and low-cost MMR methods using wearable devices.The research is conducted around three aspects as follows.(1)A systematic study on the impact of key parameters in MMR systems on recognition results is presented.As the most critical parameters in MMR systems,the sensor type,sensor sampling rate,and segmentation length used in recognition not only directly determine the integrity of input information,but also the system energy consumption.In this work,a systematical analysis of these key parameters in MMR is conducted based on large datasets and machine learning methods.Understanding the impact of these parameters on recognition results is helpful to select the optimal parameter configuration for different MMR scenarios,which is the basis to realize robust MMR systems.(2)A high-accuracy and low-cost human daily motion mode recognition(DMMR)method is presented based on its application scenario,and a DMMR system is realized.Since the main service object of DMMR is the public,from the perspective of portability and intelligent software design,the optimal implementation platform for DMMR is smartphones.Therefore,this work uses the smartphone as the platform and proposes a novel DMMR method,LCAMMR.LCAMMR firstly cuts and compresses traditional deep neural networks based on lightweight deep convolution.After that,LCAMMR designs a novel dynamic feature extraction method according to the process of time-series data processing in DMMR.This method can effectively avoid the redundant calculation of LCAMMR during continuous operation,thereby reducing the energy consumption of the entire model.Experiment results show that,compared with other methods,LCAMMR can reduce energy consumption by hundreds of times while maintaining high recognition accuracy.This makes it more suitable for mobile devices with limited computing resources and battery capacity.(3)A cascade and parallel multi-state human fall detection(FD)method is presented,and an FD system is realized.Different from DMMR,the service target of FD is the elderly,and its application scenarios mainly include nursing homes and hospitals.Therefore,FD emphasizes more on the stability and accuracy of the system.In this case,small waist-mounted customer integrated devices are the most commonly seen in FD.Therefore,this work uses a small waist-mount device as the platform and proposes a new FD method,CMFALL.The method simultaneously uses the unique signal characteristics of the human body during falls,and the modeling capability of the convolutional neural network.CMFALL firstly recognizes the fall behavior based on the characteristics of sensor signal changes during the fall and then uses the lightweight self-attention convolutional neural network proposed to balance its accuracy and energy consumption.Experiment results show that,compared with other methods,CMFALL can reduce energy consumption by thousands of times while maintaining high recognition accuracy.This makes it more suitable for low-cost wearable devices with limited computing resources and battery capacity.