Improving Indoor Localization Using Mobile UWB Sensor and Deep Reinforcement Learning

Indoor localization has posed a significant challenge for researchers. Current methodologies predominantly rely on ultra-wideband (UWB) technology and selection of an appropriate number of anchor points to achieve precise accuracy at the centimeter level. However, the efficacy of these approaches can be compromised when anchor points are incorrectly positioned due to factors such as multi-path effects. Such misalignment can lead to signal power attenuation, thereby diminishing the overall accuracy of localization. In this paper, we propose a novel solution
to address this issue. Our approach involves the utilization of deep reinforcement learning (DRL) to train a mobile UWB sensor in the identification of suitable anchor points. By leveraging DRL, we aim to mitigate the loss of transmitted signal power associated with unsuitable anchor placement. Subsequently, we conduct an evaluation to compare the performance of intelligently selected anchor points against two alternative strategies: anchor points selected with predefined constant positions and those chosen randomly. We employ the convolutional neural network (CNN) algorithm for this comparative analysis. Specifically, we utilize the received UWB signal time vector as input and predict the 2D target position using a CNN regressor to estimate the target location. Our simulation results demonstrate a significant improvement in localization accuracy when employing the DRL approach for anchor point selection. Specifically, the mean absolute error (MAE) achieved is approximately 0.09 m which represents a significant improvement compared to manual or random selection of anchor points, which provide MAEs of about 0.45 m and 1.20 m, respectively.