Session C5: Navigation and Positioning

High-Precision GNSS Augmented by data-driven NLOS Detection with Fisheye Cameras
Jianghui Geng, Innovation Academy for Precision Measurement Science and Technology, Chinese Academy of Sciences; Feng Wang, GNSS Research Center, Wuhan University
Date/Time: Friday, Sep. 20, 11:48 a.m.

The high-precision positioning of Global navigation satellite system (GNSS) paves the way toward to some emerging applications such as autonomous driving. However, the GNSS positioning performance would suffer greatly from non-line-of-sight (NLOS) signals in complex urban environments, which may introduce ranging errors of tens of meters or more. To overcome the negative impacts of NLOS, some data processing strategies based on signal characteristics or multi-sensor fusion have been proposed. But these approaches are not so effective when NLOS signals dominate GNSS observations. The fisheye camera can achieve effective NLOS detection, because it can observe the surroundings above the horizontal plane and be regarded as the eye of the GNSS receiver. In the conventional NLOS detection by using the fisheye camera, the sky view images captured by the fisheye camera are firstly segmented into sky areas and non-sky areas, and only the satellite projected to the sky areas is LOS. However, it may be problematic since NLOS can be misidentified in the cases of satellites obstructed by trees. This misjudgment would degrade or even disable GNSS in urban environments. For example, there are often landscape trees on urban roadsides, which may completely obstruct the sky. According to the conventional NLOS detection by the fisheye camera, satellites penetrating trees are considered NLOS, rendering no available satellites and GNSS solutions in these areas.
Therefore, to address this, we develop an intelligent and effective NLOS detection method and GNSS measurement weighting model based on the fisheye camera and multi-sensor fusion, aiming at robust high-precision GNSS positioning in complex urban environments. Firstly, we use deep neural networks to segment different ground objects and sky in fisheye images, achieving intelligent acquisition of environmental semantics. Then, the satellites can be projected to the segmented images and divided into LOS and obstructed satellites. Different weighting models are designed for different types of satellites to optimize the random model of satellite measurements. For LOS satellites, the conventional weighting function based on elevation angle and C/N0 will be adapted. For the satellites obstructed by buildings, they are identified as NLOS satellites, so they will be excluded in subsequent data processing. For the satellites obstructed by trees, they will not be considered as NLOS satellites. We will fuse other sensors, such as inertial sensors, to determine the corresponding weighting model. Specifically, a priori residuals of the satellite’s measurements will be calculated with the assistance of multiple sensors, which is more reliable than GNSS-only. Afterwards, the combined weighting model, which integrates the elevation angle and C/N0 with a priori residuals of the satellite’s measurements, will be implemented.
Finally, we conducted experiments to validate the advanced positioning performance of the proposed system. We found that these signals penetrating through trees have different characteristics from NLOS signals. The results show that the proposed method exhibits more effective NLOS detection than the conventional NLOS identification with fisheye cameras. The positioning precision in east, north and up components can achieve 4.2, 3.1 and 6.9 cm by using the proposed method, exhibiting improvements of about 47%, 72% and 42% in the three components, respectively, compared to solutions by using the conventional NLOS detection method. The maximum of positioning errors is decreased from several meters to only three decimeters. The rate of 3D positioning errors less than 0.3 m is also increased from below 90% to more than 99%. We, therefore, demonstrate that the proposed method outperforms the conventional NLOS detection by the fisheye camera, and the signals penetrating through trees can benefit GNSS high-precision positioning in urban environments.