overview
The full name of GNSS is Global Navigation Satellite System (Global Navigation Satellite System), which means all satellite navigation systems, including global, regional and enhanced satellite navigation systems, such as GPS in the United States, GLONASS in Russia, Galileo in Europe, Beidou satellite navigation system in China, and related enhancement systems. GNSS is widely used in outdoor robot positioning, automatic driving and other fields. Here we mainly introduce two modeling methods for the fusion of GNSS and cameras, IMUs and other sensors.
We all know that positioning and navigation based on relative positioning or recursive methods, such as inertial navigation, visual SLAM, VIO and other systems, will drift greatly as the running time increases. The GNSS positioning is an absolute positioning, and its position is not related to the previous position, so there is no drift effect. However, its positioning accuracy is very limited, and the general GPS positioning accuracy is at the meter level. Some scholars consider the fusion scheme of VIO+GNSS to realize the complementarity of the two methods, which can not only achieve global drift-free positioning, but also obtain higher accuracy in local trajectories. Here we introduce two different approaches to GNSS modeling.
The first method: direct observation
This modeling method is the most commonly used method, and the VIO+GPS scheme in most engineering implementations and papers uses GNSS direct observation modeling. Direct observation means directly using the position information output by GNSS to align with the local trajectory output by local positioning algorithms (such as VO, VIO). That is to construct the following two observation equations:
Second Approach: Degraded Observation
Degraded observations mean using GNSS raw measurements: pseudorange and Doppler shift. Pseudo-range means the distance from the GPS observation station to the satellite. Since the influence of "satellite clock and receiver clock synchronization error" has not been corrected, the measured distance includes the clock error factor, so it is called "pseudo-range". Compared with the previous method, this modeling method is more tightly fused with sensors such as cameras, IMUs, etc. When the number of tracking satellites is less than 4 and the GPS position output is unreliable, this method can still output stably.
Comparison of the effects of the two methods:
Let's compare the positioning effects of the two modeling methods in practice. Both methods use the configuration of VIO+GPS. We take VINS_Fusion[1] as an example for direct observation modeling, and GVINS[2] for degraded observation modeling. Both methods are the work of Hong Kong University of Science and Technology and are open source.
From the results, it can be seen that both methods have well eliminated the trajectory drift of VIO for long-running. Direct observations are added to GPS observations in a loosely coupled manner, while degraded observations are fused from the sensor data level in a tightly coupled manner. In comparison, the degraded observation model is more complex, taking into account more parameters and constraints (pseudorange constraints and Doppler constraints), so it can achieve higher accuracy and more robust systems. In the paper, an absolute trajectory error of 4.5m is reported in a driving distance of 5.9km. The positioning accuracy is comparable to RTK, but it is more robust than RTK because it can handle the situation when there is no satellite signal incomplete. For the method of direct observation, the model is relatively simple. From the perspective of longer-term operation (the covariance of VIO exceeds the covariance of GPS), the positioning error of the system will converge to the level of GPS.
From the perspective of optimization algorithm, the two methods are the difference between loose coupling and tight coupling. Direct observation starts from the perspective of pose graph optimization, and degraded observation starts from the perspective of sensor factor graph optimization. A globally consistent estimate can be obtained. It is worth noting that both require strict time alignment. Because satellite-based positioning has very strict requirements on time synchronization.