| In pedestrian positioning systems,a single inertial navigation system or wireless positioning system cannot guarantee both accuracy and continuity of positioning.Therefore,fusion positioning systems based on the complementary advantages of multiple positioning systems are more widely used in practical scenarios.At present,the mainstream fusion positioning systems are mainly based on filter fusion of traditional algorithms such as inertial navigation and trilateration positioning.Due to the need for a large number of empirical model parameter settings,such systems are vulnerable to modeling parameter errors of devices and environments.In order to improve the accuracy of pedestrian localization modeling for high-complexity systems,more research has started to use deep learning algorithms in recent years.However,current research has only improved the modeling accuracy of some components in the localization system by deep learning,and thus has never been able to avoid the errors caused by manual parameter settings.In this regard,this thesis proposes a fusion localization system based on an end-to-end structural deep learning model.An unsupervised transfer learning method is also proposed for the problem of pedestrian and device heterogeneity in the localization scenario.The main innovations and work of this thesis are as follows:1.We proposed SmartFPS,a wireless inertial guidance fusion localization algorithm based on end-to-end deep learning architecture,to solve the problem of introducing errors in parameter settings in filtered fusion localization systems based on empirical models.Among them: 1)the middle layer of the sub-localization network is used as the encoder output to protect feature integrity;2)an attention mechanism is introduced to improve model convergence in large scenarios;3)a recurrent neural network fuses pedestrian orientation history information;and 4)a multi-task learning approach is used for model training.Simulation validation based on log distance path loss model and Lambertian model shows that this algorithm is generalized to Wi Fi,Bluetooth,visible light and other wireless localization systems.Compared with filter fusion,SmartFPS can improve the interference suppression performance by 85.7%;compare with MM-Loc output layer fusion method,the localization accuracy can be improved by 90.5%;compare with other network decoder schemes,this scheme can improve the localization accuracy by at least 28.6%.2.For the first time,we proposed SSigGAN,a transfer learning algorithm for wireless inertial guidance fusion system based on generative adversarial network,which improves the antiinterference capability of the fusion algorithm for pedestrian and device heterogeneity.In which: 1)a transfer learning model based on generative adversarial network is used for each sub-localization network;2)the input strategy of the target domain training set is optimized and a real-time migration method based on wireless localization assistance is proposed.Simulation validation based on the log distance path loss model shows that SSigGAN can improve the fusion localization accuracy by 45.0%under the interference difference of the target domain data set.In addition,the simulation validation based on visible light localization shows that SSigGAN can be applied to the wireless localization system of Lambertian model.3.Based on the above innovations,a Bluetooth inertial fusion pedestrian localization system DBI-Loc is implemented,where 1)the algorithm uses SmartFPS and utilizes the source domain dataset for model training,and 2)the distribution alignment for pedestrian and device heterogeneity is achieved using SSigGAN with target domain data.The system validation results show that the DBILoc with SmartFPS can improve 36.5% in localization accuracy compared to filter fusion such as EKF and PF.In addition,the DBI-Loc localization accuracy can be improved by 45.8% compared to the localization system using MM-Loc algorithm,which is also an end-to-end neural network fusion system.In terms of model tranferring,the system validation results show that SSigGAN can improve the estimation accuracy of step and rotation angle of the inertial navigation subsystem by 53.3%,Bluetooth localization accuracy by 33.4%,and the overall accuracy of DBI-Loc by 31.6%.In addition,the transfer learning method using wireless localization assistance can improve the transferring performance by 37.8% compared with the random sampling of source and target domain training sets.Compared with other deep learning fusion localization systems,DBI-Loc shows significant improvement in localization accuracy.In terms of system latency,the computational latency of DBILoc is not significantly different from that of the filter-fused localization system.Overall,the DBI-Loc system based on SmartFPS and SSigGAN can meet the real-time localization in complex indoor environments and motion conditions,and achieve 0.575 m localization accuracy for different pedestrians and devices without additional data tag collection. |
