A Multi-target Device-free Localization Approach Based On Deep Learning For UHF RFID

In recent years,with the development of the Internet of Things technology,the demands of obtaining target information,especially indoor target location information,have become urgent requirements.Thus,indoor localization technology has increased its research significance and value.As the Device-Free Localization does not require the target to carry any electronic devices,its application limitations are eliminated.It has a wide range of demands in many practical scenarios and has become a hot topic in the indoor localization research fields.RFID localization has always been an important research field of indoor localization.Passive ultra-high frequency(UHF)RFID tags have been widely used in indoor localization application scenarios,due to their advantages such as small size,low cost,and powerlessness.At present,under the impact of complex channel environment,multi-path effect,severe interference of non-line-of-sight propagation,and other factors,it is still a challenging subject to implement passive localization applications with high precision and strong anti-interference capacity.In addition,when multiple undetermined targets exist in the localization scene simultaneously,it is required to overcome the difficulty of distinguishing and locating multiple targets.In order to solve these problems,this dissertation chooses Device-Free Localization for RFID as the research topic.First,the localization scene is divided into uniformly distributed grids,and the location of the target to be localized is associated with the grid and labeled.Thus,the localization problem is transformed into a classification problem.At the same time,the fingerprinting method is used to eliminate the impact of multi-path and non-line-of-sight.Afterward,a single target fingerprint database is established,based on which a deep neural network(DNN)is designed and implemented.Eventually,a Softmax classifier was used to locate the target.For the purpose of reducing the number of classes used for classification,improving prediction accuracy and computational efficiency,this dissertation adopts a two-dimensional classification approach.The simulation and experiment results demonstrated that,compared with the traditional classification algorithms such as KNN,decision tree,and random forest,the single target localization error can be limited to 0.8m with a probability of 99.75%,and the multi-target(the number of targets is 3)localization error can be limited to 1.5m with a probability of 82.1%.In order to remove the limitation on the accuracy of the two-dimensional classification algorithm brought by the grid size,this dissertation used generalized regression neural network(GRNN)algorithm to alleviate the grid size restrictions on localization accuracy.The simulation and experiment results demonstrated that the localization error can be limited to 1.5m with the probability of 84% for multi-target localization.The algorithm performance is further improved.