TL;DR
This paper introduces a novel, cost-effective system for land vehicle positioning that fuses data from low-cost sensors like MEMS inertial sensors and GPS. It tackles the challenge of maintaining accuracy during GPS outages by employing a unique distributed-dual-H∞ filtering (DDHF) mechanism and a generalized regression neural network (GRNN) for error compensation, demonstrating superior performance compared to existing methods.
What Problem Does This Paper Address?
Accurate and reliable positioning is crucial for land vehicles, especially for advanced driver-assistance systems and autonomous driving. Traditional approaches often combine Global Positioning Systems (GPS) with Inertial Navigation Systems (INS). While GPS offers good accuracy in open areas, its signals are easily blocked in urban canyons, tunnels, or under dense foliage, leading to positioning interruptions. INS, on the other hand, is self-contained but suffers from accumulating errors over time due to sensor drift, particularly in low-cost MEMS inertial sensors.
Existing solutions often rely on Kalman filters (KFs), which require accurate models of sensor errors. However, the nonlinear drift of MEMS sensors is difficult to model precisely, leading to performance degradation during GPS outages. This paper addresses the need for a robust and cost-efficient positioning system that can maintain high accuracy even when GPS signals are unavailable, specifically by overcoming the challenges posed by uncertain MEMS sensor errors.
Methods
The proposed system integrates a Reduced Inertial Sensor System (RISS) consisting of a single vertical gyroscope and two accelerometers (1G2A), low-cost GPS, and supplementary sensors like an electronic compass, wheel speed sensor, and velocity constraints.
Key methodological contributions include:
- H∞ Filtering for Attitude Estimation: To mitigate the uncertain nonlinear drift of MEMS inertial sensors, an H∞ filter is used to accurately estimate pitch and roll angles based on a vehicle kinematic model. This approach is robust to sensor noise and drift.
- Distributed-Dual-H∞ Filtering (DDHF): This novel mechanism fuses information from GPS and supplementary sensors. It comprises a Main H∞ Filter (MHF) that uses GPS data and an Auxiliary H∞ Filter (AHF) that relies solely on supplementary sensors. The difference in state vectors between MHF and AHF during GPS availability serves as crucial training data for the GRNN.
- Generalized Regression Neural Network (GRNN) for Error Compensation: A GRNN is employed to compensate for RISS position errors during GPS outages. It takes inputs such as time, RISS measurements (yaw rate, accelerations, pitch, roll), and outputs the difference between MHF and AHF state vectors. This hybrid approach leverages both learned patterns and available sensor data.
The system operates in two modes:
- Learning Mode (GPS Available): MHF and AHF run in parallel. The difference between their states is used to train the GRNN.
- Prediction Mode (GPS Outage): MHF is disabled. The trained GRNN, using AHF outputs and current sensor inputs, predicts and compensates for RISS position errors.
Datasets
The paper describes experimental validation using real-world road-test data collected from a test vehicle equipped with the described sensors. Two trajectories were used:
- Trajectory 1: Conducted in a suburban area with simulated 50-second GPS outages. This trajectory included various driving maneuvers and was used to evaluate the H∞ filter for attitude estimation and the overall positioning solution.
- Trajectory 2: An urban scenario with real GPS outages caused by infrastructure and overpasses. This trajectory aimed to further validate the system's performance in more challenging conditions.
The authors also used a high-accuracy NovAtel SPAN-CPT system as a reference for quantitative comparison.
Key Insights and Ideas
- H∞ Filter's Robustness: The H∞ filter proves highly effective in mitigating the impact of uncertain nonlinear drift in MEMS inertial sensors for pitch and roll angle estimation, yielding significantly smoother and more accurate results than direct calculation.
- DDHF for Sensor Fusion: The DDHF mechanism cleverly leverages the complementary nature of GPS and supplementary sensors. By using the difference between MHF and AHF states, it creates a valuable training signal for the GRNN, even when GPS is intermittently available.
- GRNN for Outage Compensation: The GRNN's strong approximation capabilities are well-suited for learning and predicting the complex error patterns of RISS during GPS outages, providing accurate corrections when they are most needed.
- Hybrid Approach Synergy: The combination of DDHF and GRNN creates a powerful hybrid system. DDHF handles sensor fusion and provides error estimates, while GRNN learns from these estimates to predict and compensate for errors during GPS blackouts.
- Cost-Effectiveness: By utilizing low-cost MEMS sensors and GPS, the proposed system offers a highly reliable positioning solution without the prohibitive cost of high-end inertial systems.
Key Results
- Attitude Estimation: The H∞ filtering approach reduced the Standard Deviation (STD) of pitch error by 62% and roll error by 45% compared to the original calculation method.
- Positioning Accuracy (GPS Outages):
- In Trajectory 1, the proposed GRNN-DDHF solution achieved significantly lower maximum and RMS position errors during GPS outages compared to RBF-DDHF, DDHF, KF, and KF-GENERAL solutions. For instance, maximum error was improved by an average of 36% over RBF-DDHF and 39% for RMS error.
- In Trajectory 2 (urban, real outages), the GRNN-DDHF solution again outperformed other methods, demonstrating its robustness in more challenging environments.
- Robustness to Sensor Drift: When simulated biases were introduced into the inertial sensors, the DDHF-based solutions (GRNN-DDHF and RBF-DDHF) showed much smaller performance degradation compared to KF-based solutions, highlighting the advantage of H∞ filtering's insensitivity to predefined error models.
Limitations and Future Work
- Limitations:
- The paper acknowledges that the GRNN's performance is dependent on the quality of training data.
- While the system is designed for cost-efficiency, the performance is still compared against more complex systems.
- The assumption of velocity constraints might not hold perfectly during extreme maneuvers.
- Future Work:
- The authors suggest integrating more information from the vehicle's Controller Area Network (CAN) bus, which already contains data from wheel speed sensors and some inertial sensors, to further enhance positioning accuracy at a lower cost.
- Further research could explore more advanced ANN architectures or fusion techniques.
Summary updated