Executive Summary
This report includes a thorough analysis of the concrete performance difference between tightly-coupled (TC) and loosely-coupled (LC) Visual-inertial Odometry (VIO) systems in the case of short-term and long-term Global Navigation Satellite System (GNSS) outages. VIO, a technology that combines the data from a camera and Inertial Measurement Unit (IMU) to estimate the ego motion, is becoming more common with the integration of GNSS technologies for globally accurate and continuous positioning. The coupling strategy (tight or loose) determines the fusion of these sensor streams and has a strong effect on the system behavior, in case of lost absolute positioning reference from GNSS.
Our analysis, which was based on a systematic review of the given research data, concludes that tightly-coupled systems tend to exhibit better robustness and slower position drift accumulation than loosely-coupled systems, especially in the presence of extended GNSS outages. This advantage is derived from their capabilities in using raw GNSS measurements (even of a limited number of satellites) to continuously calibrate sensor errors up to the outage point and maintain a more probabilistically consistent state estimate. However, this is at the expense of increased computational complexity.
Key findings show that in short outages (less than or equal to 30 seconds), well-implemented loosely-coupled systems are able to maintain centimeter-to-decimeter level accuracy, and are able to compete with tightly-coupled counterparts. During extended periods of outage (> 30 seconds), drift characteristics of the IMU and robustness of the VIO subsystem are the dominant factors for both architectures. Nevertheless, tightly-coupled frameworks tend to handle the transition and subsequent dead reckoning phase much better and hence more graceful degradation. The quality of the IMU is the single most important hardware component of drift with high grade sensors allowing significantly longer periods of accurate navigation than low grade MEMSsensors, irrespective of the coupling strategy.
We have identified that a critical gap in the current research community is the absence of recent, peer-reviewed studies that directly compare TC and LC VIO performance, on real outdoor datasets with simulated GNSS dropouts, and also report the formal statistical significance such as p-values or confidence intervals. This reduces the potential to make conclusive and universally applicable quantitative conclusions.
1. Architectural Foundations Tightly-Coupled vs. Loosely-Coupled VIO
The main distinction between tightly-coupled and loosely-coupled VIO systems has to do with the architecture of the sensor fusion estimator. This architectural choice has significant implications on accuracy, robustness and computational requirements, particularly in GNSS degraded environments.
1.1. Loosely-Coupled Architecture (LC).
In a loosely-coupled system, the GNSS receiver and the VIO subsystem are independent modules, each giving their own state estimate (e.g. position, velocity)1. These separate solutions are then fed into a secondary fusion filter, usually a Kalman filter, which combines them to yield a final, blended navigation solution2.
- Mechanism: The VIO subsystem gives a high frequency, locally correct relative pose estimate, which has the tendency of drifting over time. The GNSS receiver is a low-frequency, globally accurate absolute position receiver, but is prone to outages. The secondary filter uses the GNSS position to correct the accumulated drift of the VIO solution3.
- Advantages: This cascaded approach is simpler to implement, and is more modular. Different VIO and GNSS components can be interchanged fairly easily. The computational load is often less than in tightly-coupled systems.
- Disadvantages: This is not an optimal architecture from an information theoretic point of view. If the GNSS receiver cannot see at least 4 satellites to compute a full PVT (Position, Velocity, Time) solution, no information from GNSS is used in the fusion filter, even if the raw measurements of one, two or three satellites are available4. This causes hard failure mode where the system reverts completely to VIO dead-reckoning, and any possible aiding from a degraded GNSS signal is lost.
1.2. Tightly-Coupled (TC) Architecture
In a tightly-coupled system, no intermediate state estimation is done. Instead, raw measurements from all sensors (delta-V and delta-theta from the IMU, visual feature tracks from the camera, and raw GNSS observables, or pseudoranges, carrier phases, and Doppler shifts) are directly fed into a single, comprehensive state estimator5. This is often implemented using an Extended Kalman Filter (EKF) or more commonly in state of the art systems using a factor graph optimization (FGO) framework6.
- Mechanism: The unified estimator forms a large state vector of the position, velocity and orientation of the platform, sensor biases and other calibration parameters and solves for these variables. By processing raw measurements it can make use of all the information at once7 8.
- Advantages: This approach is much more robust. It can continue to benefit from the use of GNSS information even when less than four satellites are visible, using measurements from a single satellite to constrain clock drift or from two or three to constrain position.This continuous aiding enables improved online calibration of IMU biases which results in reduced drift during a full outage .
- Disadvantages: Tightly-coupled systems are much more complicated to design and implement. The computational burden is usually higher, because of the larger state vector and the complexity of the joint optimization problem, which is particularly high in the framework of factor graphs, where the graph size increases with the length of the trajectory.
2. Performance During Short-Term GNSS Outages ( ≤ 30 seconds)
During short outages of the GNSS signal, both architectures will work fine, but the mechanisms and failure points are different.
2.1. Quantitative Analysis Of Drift
Numerous studies offer quantitative data for drift for outages of a few hours, but the results are variable depending on the quality of the IMUs and details about the system.
- Tightly-Coupled Systems:
- For systems operating with tactical grade IMUs, a horizontal position drift of about 0.3 meters during a 30 second partial GNSS outage is a commonly reported number.
- With high-grade IMUs, it’s even better with drift being reported as almost linear, < 0.3 meters in 10 seconds, and 1 to 2 meters in 30 seconds.
- Other tests on tightly-coupled INS/GNSS systems give position errors of 0.20 meters after 10 seconds and 0.55 meters after 30 seconds. One study indicated an erroneous position of about 25 meters after only a 15 second outage, indicating the high variability that depends on test conditions and sensor quality.
- Loosely-Coupled Systems:
A loosely-coupled RTK/Vision integration system was demonstrated to attain centimeter level positioning for about the first 30 seconds of a GPS signal outage . This means that if the VIO component is much more accurate and stable, then the LC approach can be used to effectively bridge short gaps.
2.2. Comparative Insights
In the short term regime, the difference in performance is often subtle. A high quality VIO subsystem can enable an LC system to coast with minimal drift, equaling the performance of a TC system. The main benefit of the TC system is its resilience to degraded, rather than total absence, of the GNSS signals. It can use data from the remaining visible satellites up until a complete outage occurs to ensure that its state estimate (and especially its IMU bias estimates) is as accurate as possible before it enters the dead reckoning phase.
3. Performance Under Long lasting GNSS Outages (> 30 seconds)
As the outage of the GNSS is prolonged, the inherent drift of the dead reckoning sensors (IMU and VIO) becomes the main cause of error and the architectural differences between TC and LC systems become more evident.
3.1. Quantitative Analysis Of Drift
- Tightly-Coupled Systems:
- Performance continues to deteriorate over time. One study reported position accuracy degradation of 2.5 m (1-sigma) for 30-second outages and < 6 m (1-sigma) for 60-second outages.
- In a more optimistic test, a system’s position accuracy is reported to have remained within nominal GNSS accuracy for more than 2.5 minutes, with a mean position drift of less than half a meter following a 50 second outage.
- For VIO-centric tightly-coupled systems (without GNSS aiding), maximum position drifts over an entire trajectory has been measured at 3.4-7.9 meters demonstrating the baseline drift that GNSS is supposed to help correct.
- Loosely-Coupled Systems:
- One of the standouts for a loosely-coupled RTK/Vision system was that it gave a position error of < 0.8 meters during a 200-second (3.3-minute) outage. Furthermore, a dynamic SLAM-based approach was reported to achieve a decimeter level of precision during long outages (within 10 minutes).
- In another test where a long duration outage was used, a loosely-coupled estimator showed an error standard deviation of 1 meter in both east and north directions.
3.2. Comparative Insights
In case of prolonged outages, the VIO subsystem essentially becomes the main navigation source for both architectures. The performance of the system is therefore fundamentally limited by the drift characteristics of the VIO, which are dependent on distance traveled and quality of the visual environment.
The benefit of the TC architecture in this regime is its better capability of having modeled and compensated for sensor errors, especially IMU biases, based on all the available data before the outage began. This results in a better propagation of the state estimate during dead reckoning and generally an accumulation of error that is slower than the rate in which it would occur without it. The LC system having relied on filtered position corrections may be carrying forward unobserved errors in its VIO subsystem which can manifest as larger drift during the long outage.
4. Important Performance Influences on Drift
Coupling strategy is just one of the variables chosen. Two other factors are of utmost importance when it comes to drift behavior during GNSS outages.
4.1. Quality of Inertial Measurement Unit (IMU)
The quality of the IMU is really perhaps the most important factor. Left uncorrected, IMUs have position errors that increase cubically with time.
- Low-Grade vs. High-Grade: Low-grade, consumer-grade MEMS IMUs have excessive noise and erratic biases, which leads to high drift.Lower-grade (e.g., tactical or navigation-grade) IMUs have much less noise and more stable biases, so they can stay accurate for much longer periods of time during an outage.A navigation-grade IMU can effectively beat a MEMS-based IMU by orders of magnitude over a 20-minute black-out period.
- Impact on Coupling: Although a better IMU has a positive effect on both TC and LC systems, the effect is more significant in a TC system. The excellent observability of the TC filter makes it possible to estimate more precisely the subtle biases of a high-grade IMU to maximize its performance. Even for a low grade IMU, the drift can be so fast that a TC estimator is quickly overwhelmed by it.
4.2. Visual Environment Robustness (VERB)
During a GNSS outage, VIO is of key importance. The VIO algorithm’s performance is very sensitive to the environment in which it is being run. For example, under bad texture conditions, dynamic light conditions, or fast moving scenarios, VIO might fail or build up drift quickly. The second is a failure common to both LC and TC systems as it is a failure of their common dead-reckoning sensor.
5. Computational Load and Real Time Issues
An important trade-off between the two architectures is the performance vs. the computational cost.
- In general, loosely-coupled systems are computationally cheaper than highly-coupled systems. Their modular cascaded filter architecture is easier and more efficient to implement.
- Tightly-Coupled Systems (in particular, those based on factor graph optimization) are more costly. The optimization problem becomes larger over time and it is difficult to solve it in real time.
- Hardware Acceleration: Due to the computational requirements of TC systems, hardware acceleration through GPUs or FPGAs is also being investigated. GPU-accelerated VIO front-ends have been demonstrated to have very high throughput (e.g. <~200 fps), allowing for real-time operation on embedded platforms9. However, because of a long GNSS outage, the GNSS-related calculations are not carried out, but the VIO and IMU processing along with the optimization graph growing is the primary computation factor. However, FPGAs may be a viable solution for certain applications because memory scalability is a bottleneck in utilizing GPUs for real-time graph processing.
6. State of Benchmark and Evaluate
Critical, comparative assessment is a crucial factor in progressing the field. However, the current ecosystem does have some distinct limitations.
- Existing Datasets: Although there are well-known VIO datasets such as EuRoC and TUM-VI, these do not contain GNSS data. . More recent datasets have been published to fill this gap, for use in GNSS-VIO fusion research. These include the GVINS-Dataset from HKUST and the AaltoVision/vio-gnss-dataset which contain synchronized visual, inertial and raw GNSS measurements.
- Research Gap: A consistent trend across the data analysis is that recent (2022-2024) peer-reviewed publications are not available which directly, side-by-side compare the TC and LC VIO on outdoor datasets with simulated configurable GNSS outages and report formal statistical significance testing (i.e. p-values, confidence intervals) on the error metrics. While a majority of the works reports on results from their own proposed system and compares it to a baseline, robust statistically-supported comparisons of coupling strategies under the same conditions are scarce. This makes it hard to generalize the quantitative results included in this report.
7. Conclusion
Thus, a fundamental trade-off between robustness and complexity exists between tightly-coupled and a loosely-coupled VIO architecture.
- Tightly-coupled VIO shows the highest theoretical and practical robustness especially during the beginning of and within long GNSS outages. Its in-line capability to use all raw sensor information including degraded signal allows for better sensor calibration and more graceful degradation of performance.
- Loosely coupled VIO is low in computational complexity. In the case of short intermittent GNSS outages and a very robust VIO subsystem, it can perform competitively.
Finally, the position drift during a GNSS outage is also a function of the entire system. The most influential hardware factor is the quality of the IMU while the performance of the VIO algorithm in the particular visual environment is the most influential software factor. While a tightly-coupled architecture gives the most solid structure for combining these elements, its performance will still be fundamentally limited by the quality of its inputs. A glaring lack of statistically significant comparative studies is still a major gap and the future scientific investigation needs to be along such rigorous assessment to provide concrete guide for system designer.
References
- GitHub – aaronboda24/Loose-GNSS-IMU: Loosely coupled integration of GNSS and IMU, 2018
- Victor Canepa and Pablo Servidia, “Design and evaluation of INS/GNSS loose and tight coupling applied to launch vehicle integrated navigation,” P. 5, Metascience in Aerospace, 1(1): 66–109. doi: 10.3934/mina.2024004
- Hujun Bao, Weijian Xie, Quanhao Qian, Danpeng Chen, Shangjin Zhai, Nan Wang, Guofeng Zhang, “Robust Tightly-Coupled Visual-Inertial Odometry with Pre-built Maps in High Latency Situations,” IEEE Transactions on Visualization and Computer Graphics, Volume: 28, Issue: 5, 2022, P. 2 DOI: 10.1109/TVCG.2022.3150495
- Design and evaluation of INS/GNSS loose and tight coupling applied to launch vehicle integrated navigation Page30 Figure 16.
- Loosely Coupled & Tightly Coupled INS & GNSS [2024 Guide]
- Factor Graph Fusion Localization Method with Tight and Loose Coupling of GNSS/IMU and Odometry Page1
- VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator Page 1
- Tightly-coupled Fusion of Global Positional Measurements in Optimization-based Visual-Inertial Odometry Page 6
- Faster than FAST: GPU-Accelerated Frontend for High-Speed VIO Page1&7
