Research On The Technology Of Passive Human Localization Base On Pyroelectric Infrared Signal

Currently,the information of human locations becomes important in more and more applications,e.g.smart city,smart home,and smart security.The techniques of human localization are booming and each of them usually has its own suitable application scenarios.Recently,due to the advantages of low power consumption,privacy protection,and low-light adaptation,the localization technique based on the pyroelectric infrared(PIR)sensor is attracting more and more attention from both academic and industrial fields.However,existing PIR-based localization techniques still have much room to be improved.For example,most practical applications require sub-meter level human localization accuracy.To meet this requirement,existing PIR-based localization techniques usually require high deployment density of the PIR sensor.Therefore,to reduce the deployment density of the sub-metter level PIR-based localization system,this dissertation deeply explores the analog signal of PIR sensors to extract information of human locations,rather than inferring locations based on the binary information extracted from PIR sensors' analog signal.Specifically,this dissertation proposes a series of methods that can achieve sub-metter level PIR-based localization.The main contributed of this dissertation are as following:(1)This dissertation systematacially build the relationship between the human position and PIR sensor's output for the first time.Specifically,by integrating the model of PIR sensors,the model of infra radiation,and the model of the Fresnel lens array,this dissertation quantificationally describe the relationship between the PIR output and the position of a human.Experiments shows that the mean difference between the prediction value of the axis of a PIR sensor's detection zone and the measured value is only about 0.5 degree.In addition,this dissertation also proposes a filter based on the model of a PIR sensor,which can alleviate the a PIR sensor's signal distortion.(2)Based on the character of PIR signal that refracted by the Fresnel lens array,this dissertation proposes a novel localization method PIRATES based on the information of azimuth change,which is much different from the traditional localization method like trilateral positioning.Specifically,this dissertation builds the sensing model of a PIR sensor with Fresnel lens array.Based on the above model,this dissertation proposes a method that inversely filters the signal caused by the human motion when the human moves across the detection zones of the Fresnel lens array.After obtaining the human's azimuth change,the human's location can be inferred based on the azimuth change.Compared with existing PIRbased localization method that utilizes the PIR sensor's analog signal,PIRATES improves the localization accuracy by 24% and reduce the deployment density by 50%.(3)To further improve the multi-person localization accuracy under the constraint of low deployment density,this dissertation proposes another PIR-based localization method PIRNet that is based on deep learning.Specifically,to improve the neural network's generalization ability in the PIR-based localization task,this dissertation proposes to design the neural network based on the idea of modular designing.Moreover,to further improve the generalization ability,this dissertation also proposes a series of data augmentation methods based on the physical character of the PIR sensor.Compared with PIRATES,with a same deployment density,the mean localization error of PIRNet is about 0.4m lower in the multiperson scenarios.Compared with traditional method based on binary preprocessing,PIRNet can obtain the similar localization accuracy,while the deployment density of PIRNet is about80% lower than the traditional methods.(4)To decrease the training cost of the PIR localization method based on deep learning,this dissertation proposes another method Deep PIRATES that combines deep learning and the geometric localization method.Different from PIRNet,Deep PIRATES does not utilizes the end-to-end neural network architecture.Through Deep PIRATES,we can obtain the localization accuracy similar to PIRNet and the model can be applied in scenarios of any kind of deployment strategies without retraining.The above methods proposed in this dissertation significantly improves the practically of PIR-based localization system and makes it promising in some specific applications.In the future,we will go on to improve the convenience of PIR-based localization,for example the methods of single-point deployment and overlook deployment.