Research And Implementation Of Multi-person Behavior Sensing Technology Based On Millimeter Wave

In recent years,innovative achievements in the field of human behavior sensing technology have brought tremendous convenience to people’s lives in areas such as health and entertainment.Human behavior sensing technology based on wireless radio frequency signals has overcome the shortcomings of visual or wearable device-based solutions in terms of privacy leakage and invasive sensing.Among them,millimeterwave technology has become a hot research area due to its advantages in privacy security and fine-grained sensing.However,most of the existing research is carried out in single-person scenarios,while actual application scenarios are often complex multiperson interaction scenarios.In order to achieve human behavior sensing in multi-person scenarios,two key problems need to be solved:(1)how to accurately cluster and track multiple targets in the sensing scene;(2)how to design robust behavior sensing algorithms in localization and cloud application scenarios.This paper proposes three key designs to address the above challenges:(1)applying group tracking algorithms to accurately track and cluster multiple targets and complete point cloud denoising;(2)a rule-based behavior sensing algorithm that uses 27 feature values extracted from target data and point cloud data,combined with a finite state machine model,to achieve precise sensing on radar hardware;(3)the behavior sensing neural network Multi-HAR,which uses 3D-CNN to learn the spatial features of point clouds and Bi-LSTM to learn the temporal features of point clouds,to achieve behavior classification.This paper implements a human behavior sensing system that integrates clustering tracking,behavior sensing,OTA upgrading,and data transmission in multi-person scenarios,which can sense five daily behaviors including standing,walking,sitting,bending,and falling.In this system,this paper conducts accuracy,complexity,and stability tests on the rule-based algorithm.The results show that the algorithm achieves an average accuracy of 83.38%with a response time of 8ms and a memory requirement of 9KB.At the same time,using behavior datasets to train and evaluate the Multi-HAR model,the results show that the model achieves an average accuracy of 85.76%,which is 11.4%higher than the Transformer-based MM-HAT model in multi-person scenarios.Through experimental verification,this system and algorithm have the advantages of strong stability,high accuracy,and fast response speed in multi-person scenarios.