Session B5: Receiver Design, Signal Processing, and Antennas

Detection and Mitigation of NLOS and Multipath Effects by Multi-Task Learning
Ellarizza Fredeluces and Nobuaki Kubo, Tokyo University of Marine Science and Technology
Location: Beacon B

The continuous rise of autonomous vehicle (AV) technologies requires high sensor accuracy to ensure the safety and efficiency of AV operations. One of the de facto sensors is Global Navigation Satellite Systems (GNSS) receiver, and it is often integrated by inertial measurement unit (IMU), LiDAR, and cameras to enhance the overall performance of AV operations. Although sensor fusion promises better results, it is still of utmost importance to address the common errors in each sensor. When done, this would result further improvement to the integrated results, thereby also improving the AV operations. Here, we focus on addressing the multipath errors of a GNSS receiver.
In challenging environment like urban canyons, GNSS signals is affected by nearby objects such as tall buildings, walls, vehicles, and ground. In an ideal scenario like open sky environment, direct line-of-sight (LOS) signal is clearly viewed and received by the GNSS receiver. However, presence of obstructions in urban canyons causes the signals to be blocked, diffracted or reflected. LOS signals are limited due to signal blockage making a poor satellite geomery, and reception of reflected signals also occurs especially for high sensitive receivers. Both scenarios lead to very much degraded positioning accuracy. Reflected signals include two phenomena: non-line-of-sight (NLOS) signals and multipath signals. NLOS signal contains only the reflected signal because the direct signal is blocked, while multipath signal contains both the direct and reflected signals. Both phenomena significantly introduced large errors, typically tens of meters, in the pseudorange measurement due to path delay. In the signal processing domain, NLOS signal reception and multipath interference distort the code correlation waveform. For instance in GPS L1C/A, a clear and unique correlation peak is seen corresponding to the presence of direct path of signal. When multipath interference occurs, the correlation waveform will contain 2 peaks, one for the direct path and one for reflected path (assuming only one multipath signal is present). The multipath signal peak is at a later timing due to it being delayed by the reflection. In the case of NLOS signal reception, it is similar to multipath intereference except that peak for the direct signal path is not present. Its correlation waveform is heavily distorted by both signal reflection and diffraction. These characteristics in the correlation waveforms gives us the ability to address NLOS and multipath effects as early as possible in the GNSS receiver chain.
To be able to get best performance of GNSS positioning in urban environment, NLOS and multipath effects should be minimized as much as possible. In this study, we propose an approach to detect and mitigate the effect of NLOS and multipath using the multi-correlator outputs as inputs to a neural network model. The trained model will perform both classification and regression tasks. There have been a growing number of literatures that show the promising effectiveness of addressing the NLOS and multipath interference using artificial intelligence (AI) tools both in simulated and real environment. However, there is still a gap to fill as the current trend of these researches focus only one of the two tasks: detection or mitigation. In addition, only a few research actually evaluated the positioning performance once the trained models have been applied. Thus, the contribution of this study is that we leverage the ability of AI to address the NLOS and multipath interference, by first classifying the signal as LOS or NLOS, and then estimating the multipath parameters to mitigate their errors. As a result this would give us the ability to apply proper handling to NLOS signal, like fully excluding their measurements in the position calculation or applying weighting scheme, and estimate the multipath errors.
We selected multiple sites in Tokyo, Japan which are surrounded by high-rise buildings. We collected intermediate frequency (IF) data for each sites using GTEC RFFE front-end by TeleOrbit and post-processed them later using our in-house software receiver. We only use L1 frequency band. To generate the training datasets, multi-correlator outputs from each satellites were calculated. For the signal classification part, we used the ge-gnss-visibility tool which is available is GitHub to label LOS and NLOS satellites. This tool generates virtual fisheye zenith images from Google Earth at a certain location and automatically determines satellite visibility [1]. For the estimation of multipath parameter, we used an antenna precise position to calculate pseudorange residuals for each satellite. This is further explained in [2]. Once the data were prepared, we then apply data augmentation to help improve the generalization of the model to be trained. We train a neural network model using PyTorch, and then deploy it to our software receiver. In evaluating the results, we check the overall performance of the model by evaluating it against various classification and regression metrics. Lastly, we evaluate the positioning accuracy when the model is used and not used both in single point positioning (SPP) and differential GNSS (DGNSS).
[1] Suzuki, T. (2024). ge-gnss-visibility. https://github.com/taroz/ge-gnss-visibility.
[2] Kubo, N., Kobayashi, K., & Furukawa, R. (2020). GNSS Multipath Detection Using Continuous Time-Series C/N0. Sensors, 20(14), 4059. https://doi.org/10.3390/s20144059