High-Accuracy Multistatic UWB Radar Positioning Using Low-Cost Devices Based on Dense Convolutional Network and a DBSCAN Denoiser

Abstract

Ultra-wideband (UWB) radar positioning plays an important role in non-cooperative personnel positioning or device-free positioning. Recent work has reformulated the time-of-flight (ToF) estimation task as a 2-D image processing problem using residual convolutional neural networks (RCNNs), avoiding intricate procedures of traditional methods. Despite its benefits, the RCNN struggles with fully exploiting feature reutilization, resulting in significant errors in ToF estimation. Although particle filters (PFs) can alleviate this problem, the constant parameters in weight estimation will affect the positioning accuracy. Therefore, we first adopt a dense convolutional network (DenseNet) to replace the RCNN to enhance the feature reutilization and improve the accuracy of ToF estimation. Additionally, we design a method for outlier and anomaly cluster elimination in the ToF time series based on density-based spatial clustering of applications with noise (DBSCAN) clustering, effectively suppressing observation noises. Finally, to address the insufficient adaptability of parameters in PF weight estimation, we improve the loss function in the DenseNet, thereby enabling it to dynamically output the variance of ToF. We verify the effectiveness and generalizability of our proposed method through an open-source dataset collected with low-cost UWB devices. Compared with a classic method, the average root mean square error (RMSE) of the proposed method within the positioning area decreases by 37.1%. Furthermore, through repeated experiments across three distinct scenarios, our method demonstrates RMSE reductions of 20.1%, 36.5%, and 28.5%, respectively, compared to an existing RCNN-based approach.