Indoor non-line-of-sight and multipath detection using deep learning approach

Abstract

Global navigation satellite systems (GNSSs) provide sufficient position accuracy outdoors but can perform poorly in indoor situations. Therefore, various techniques have been presented to meet the indoor requirements, among which the pseudolite system (PLS) is widely studied. A PLS has several error sources including clock error, thermal noise and path-dependent error; among them path-dependent errors, e.g., non-line-of-sight (NLOS, i.e., the direct ray is blocked, only reflected rays received) and multipath (i.e., the direct ray is received together with the reflected rays), are the most intractable to deal with and always cause large range errors. The indoor furniture and flat surfaces of the walls and ceilings make NLOS and multipath quite severe. Detecting NLOS and multipath is a classification problem, which can be well tackled by deep learning approaches. Deep learning algorithms are functions of nonlinear regression using thousands of parameters that distributed in many hidden layers whose values are determined by training phase. We present an NLOS and multipath detecting network (NMDN) that consists of five convolution layers and two fully connected layers; we feed NMDN with the outputs of a receiver’s tracking loop; subsequently, it tells us the detection results. The training and testing data sets are generated by a GNSS software receiver using intermediate frequency signal collected from an indoor PLS. The presented method is compared with two support vector machines, which are the traditional methods for classification, and shows an improvement of up to 45% in overall classification accuracy. The impacts of indoor NLOS and multipath are also analyzed with double difference observables. The results show NLOS causes more serious range error than multipath. The proposed method is suitable for detecting NLOS and multipath when the receiver operating environment is not stable.