| Human activity recognition(HAR)via millimeter wave(mmWave)sensing is a key to realize human-computer interaction(HCI),e.g.,activity assistance and living state monitoring.Prior work has shown the feasibility of HAR by utilizing mmWave radar,but it falls short of some real-world issues:most systems can only realize the short-range recognition but not the long-range,have poor accuracy in the noisy environment and are unable to give real-time response due to long latency.Therefore,based on mm Wave radar,the thesis studies the above problems in HAR,and gives the specific research contents as follows:1.Design a real-time long-range HAR system via mm Wave radar.The system is composed of several subsystems,the core functions of which are noise reduction processing,voxelization processing,design of lightweight HAR neural network(NN),and realization of real-time system.2.Collection of experimental data based on mmWave radar.This thesis analyzes the basic principle of mm Wave radar and its ranging and velocity measuring principle,and further deduces the calculation method to obtain characteristic values,designs and realizes the collection of experimental data required in this thesis.3.Propose an optimal clustering algorithm,called U-DBSCAN,to realize data noise reduction and preprocessing based on DBSCAN.In ideal environment,the Density Based Spatial Clustering of Applications with Noise(DBSCAN)algorithm can aggregate human-orientated movement(HOM)point clouds from noise,but in complex environment,it can not achieve.Therefore,based on DBSCAN,this thesis proposes an optimized clustering algorithm U-DBSCAN to achieve noise reduction.Finally,this thesis converts HOM point clouds after noise reduction into voxels,and integrates it into an integral temporal flow(ITF)information,which was put into NN for recognition.4.Design a lightweight neural network HARnet to solve the problem of long recognition latency caused by too deep network and too many parameters in current work.This network deals with the ITF as a whole and reduces the cost of recognition.Based on this,a real-time HAR system m-Activity is designed and implemented.The experimental results show that HARnet’s offline recognition accuracy can reach 93.25%,and the latency is less than 22ms,which meets the design requirements of lightweight.And the real-time accuracy of real-time system m-Activity reaches 89.07%on average,which has certain robustness in multiple scenes. |
