| Aided by the advancement and development of mobile sensing,human activity recognition(HAR)shows significant applicable values in a wide range of fields,like health care,smart home and intelligent surveillance.The existing wearable sensor-based HAR approaches bring users discomfort and extra burdens of charging and maintenance.Visionbased HAR approaches suffer from light conditions and always arise privacy concerns.In recent years,an increasing number of researchers have turned their attention to wireless sensing field.While millimeter-wave(mmWave)based sensing is capable to address the above problems with its advantages on fine-grained sensing ability and privacy protection.Due to the sparsity and susceptibility of mmWave radar generated point cloud,the existing researches are faced with the problem of a limited number of activities,especially for those indistinguishable activities.This thesis presents a mmWave based HAR system,which not only works well on a more varied activity set with robustness but also consumes less computational resources compared with other state-of-the-art related works.For that purpose,this project employs point cloud information as well as target information generated by the tracking algorithm.These different scales of information can complement each other and contribute to model robustness.More importantly,since the occurrence processes of most activities contain temporal information,both spatial and temporal dimension features are extracted and Transformer is employed to process temporal sequence,which enriches feature dimensions and optimizes computational efficiency.The HAR system is realized on a mainstream mmWave radar sensor.A sample throttling module and a result decision module are deployed before and after the HAR model specifically for the purpose of saving computational consumption and improving system accuracy.Seven kinds of activities are collected in total for evaluation,which includes walking,running,jumping,standing,sitting,falling,and bending.The proposed algorithm achieves an accuracy of 90.46%,which surpasses the existing work by up to 34.76%(RadHAR)and 20.05%(MMPoint-GNN). |
