Study Of Key Techniques Of Multi-person Sensing Based On Distributed Bio-radar

As a new concept radar for human target detection,bio-radar could non-contact detect changes in echo parameters caused by human activities with electromagnetic waves to obtain vital,spatial,behavioral and other information of human targets,which has a wide range of application prospects in biomedicine,public security,emergency rescue,military and other fields with its advantages of not being affected by light and penetrating detection.At the present stage,the technology of bio-radar has already been successful exploited in detection,identification,and positioning of single human target.However,with the continuous deepening of its research and application,especially the rapid development of new concepts and technologies such as enlarged health,the Internet of Things,Smart Homes,and Smart Cars,the application scenarios of bio-radar are changing from special occasions such as emergency rescue,wall-through detection,and clinical monitoring to routine occasions such as homes,public places,and vehicles.Accordingly,the application target is changing from special population such as injured individuals,terrorists,and patients to ordinary healthy population such as the elderly and drivers,resulting in a strong demand for multi-person target detection.However,the technology of multi-person targets detection based on bio-radar has become an urgent challenge at the present stage,mainly due to its detection from near to far in single-view mode using array antenna or system,which makes it difficult to deal with the randomness and uncertainty of the orientation,posture and position of multi-person targets when detecting ordinary healthy people in routine occasions.Surrounding the above-mentioned challenge,a series of related technologies based on the advantages of multi-view detection of distributed architecture are studied in this dissertation,with the aim of improving the robustness of bio-radar to human orientation,posture,position and other attributes when sensing spatial,physiological,and behavioral information of multi-person targets.The main research work and results are as follows:1.A 6-node distributed bio-radar system is developed to address the challenge of bioradar to directly integrated or transformed into a distributed system due to the limitations in size,cost,and complexity.The distributed bio-radar system is formed by 6 nodes,where every single node is equipped with 77 GHz millimeter-wave FMCW transceiver system,microstrip patch antenna array,integrated RF transceiver front-end,two-stage step-down power supply,FLASH storage,and stacked PCB design,with the range resolution up to0.051 m and speed resolution up to 0.0425 m/s.A software control method is implemented to solve the time synchronization issue in multi-view detection mode of the distributed system.The "1 master node + 5 slave nodes" network architecture is adopted in this method,with the master node as the core to manage and control the detection process of the system,achieved a time synchronization accuracy of up to 53 ms average error within the 6-node data frame.A low-cost construction method for the distributed biol-radar system is formed accordingly,which not only meets spatial,physiological,and behavioral information perception needs of multi-person targets but also offers flexibility and high-resolution in a small-sized package.2.A method for positioning and tracking of multi-person targets is studied based on the spatial information acquisition of multi-person targets in distributed architecture.The method could effectively take advantage of multi-view detection to solve the problem of random human target position,orientation,posture,state and other attributes through target positioning and tracking in single-node view with multi-node multi-person target identity association and spatial information fusion in multi-node view,based on the range-velocityangle three-dimensional point cloud features from radar echoes.And as a consequence,accurate positioning and tracking of multi-person targets in four categories and seven cases was achieved,with superior integrity than that in single-node view.More importantly,this method can be used to obtain stationary human target location and moving human target trajectory,and distinguish each human target from the spatial information level,so that the subsequent physiological detection and behavior recognition of multiple human targets can be transformed into physiological detection or behavior recognition of single human targets.More importantly,this method can separate each human target in the spatial information using the position of stationary human targets and the trajectory of moving human targets,transforming the subsequent vital signs detection and activity recognition tasks of multi-person targets into that of single human target.3.Surrounding the challenge of detecting vital signs from multi-person targets within the distributed architecture,the method for this detection task is investigated.To address the issue of stable detection of respiratory signal from stationary human targets in uncertain position,orientation,and posture,the ICA and adaptive Kalman filter were employed to enhance detection of multi-view data fusion.Compared with ICA,the Kalman filter utilizes a process model to correct the adverse effects of data in inferior views on fusion,providing a comprehensive full-node fusion method without considering input combinations.On this basis,the respiratory signal detection for two stationary human targets using Kalman filter was implemented,with the effectiveness and robustness of distributed bio-radar systems for detecting vital signs from random multi-person targets verified.4.Research into methods for activity recognition is conducted with an emphasis on behavior information from multi-person targets within the distributed architecture.Based on the multi-view human speed-range-angle 3D point cloud features obtained from distributed bio-radar,the dual attention multi-view deep convolutional neural network(DAMVDCNN)fusion recognition model is proposed,with an innovative proposal is made to use the "view attention mechanism + feature attention mechanism" to suppress the influence of disadvantaged channels and regions during distributed system detection,in response to phenomenon of "multi-view heterogeneous backlash" during multi-view fusion recognition caused by the dual layer heterogeneity of multi-view point cloud information in the "view space" and "within view point cloud feature space".The experimental results demonstrate that the network achieves 97.78% accuracy in recognizing 8 types of activities of single human target,as well as 100% accuracy for one type and 83.33% accuracy for another type when dealing with two human targets respectively,indicating the effectiveness and robustness of the distributed biol-radar system in multi-person targets activity recognition.The key innovations of this dissertation are as follows:1.A set of 6-node distributed bio-radar system composed of miniaturized and integrated FMCW transceiver sensor nodes is developed through single-node integrated transreceiving and multi-node software synchronization,which not only has reached international standards in terms of the number of nodes but addressed the issue of multi-node data synchronization through software control,offering a cost-effective solution for constructing distributed bio-radar systems with more nodes in practical applications.2.A method for comprehensive perception of multi-person targets and multiinformation based on spatial information separation with multi-view fusion enhancement is proposed,which utilizes the distance-velocity-angle three-dimensional point cloud features of human target radar echoes in the multi-view detection mode of distributed system to locate and track multi-person targets.By separating multi-person targets accurately through spatial information,the vital signs detection and activity recognition of multi-person targets is transformed into single target detection task,which would provide a new approach to improving the robustness of multi-person target detection with bio-radar technology.3.A dual attention multi-view deep convolutional neural network(DA-MVDCNN)fusion recognition model based on feature-perspective dual-layer attention mechanism for multi-view behavior fusion recognition has been proposed,aiming to solve the problem of "multi-view heterogeneity backlash" caused by the dual layer heterogeneity of multi-view point cloud information in the "view space" and "within view point cloud feature space" during fusion recognition.Based on the dual layer attention mechanism,the advantage channel and advantage point cloud space are autonomously focused to enhance the accuracy and robustness of recognition,offering a valuable method for data fusion processing of distributed bio-radars.