TL;DR
This paper introduces a novel method for robust vehicle localization using GNSS data, even in challenging urban environments with signal obstructions. By employing an "a contrario" approach to detect and remove faulty pseudo-range and Doppler measurements, their proposed Rao-Blackwellized Particle Filter (RBPF) significantly improves localization accuracy and reliability.
What Problem Does This Paper Address?
Accurate vehicle localization is crucial for autonomous systems and intelligent transportation. Global Navigation Satellite Systems (GNSS), like GPS, are widely used but struggle in urban canyons and areas with obstacles. These environments lead to Non-Line-Of-Sight (NLOS) and multipath receptions, which corrupt GNSS measurements (pseudo-range and Doppler) with faulty data. This corruption can severely degrade localization accuracy, making it unreliable. While some methods combine GNSS with other sensors or use robust estimation techniques, this paper focuses on improving GNSS-only localization by specifically addressing the issue of faulty measurements.
Methods
The core of the proposed solution lies in a two-step process:
- Outlier Detection (A Contrario Approach): The authors adapt an "a contrario" modeling technique, specifically using the Number of False Alarms (NFA) criterion. This method aims to identify measurements that are "too regular to occur by chance" under a naive model of outliers. Unlike traditional methods that require setting a threshold, this approach optimizes the NFA to find the subset of "inlier" measurements. Crucially, this method is extended to simultaneously consider both Pseudo-Range (PR) and Doppler (Dp) measurements, allowing for the detection of outliers in both data streams.
- Robust Localization (Rao-Blackwellized Particle Filter - RBPF): The detected inlier measurements are then fed into a sophisticated localization filter. The paper proposes using a Rao-Blackwellized Particle Filter (RBPF). RBPF is an advancement over standard Particle Filters (PF) because it can handle higher-dimensional state spaces more efficiently. It achieves this by splitting the state vector into a linear part (estimated analytically using an Extended Kalman Filter - EKF) and a non-linear part (approximated using particle filtering). This allows for a more accurate estimation of both the receiver's position and velocity, even with a manageable number of particles.
Datasets
The experiments were conducted using data acquired from a vehicle equipped with GPS receivers traveling in urban canyons and forest areas. The "ground truth" for localization was provided by a high-cost, multi-frequency GPS RTK (Real-Time Kinematic) system, which offers centimeter-level accuracy in fixed solutions. The paper does not introduce a new benchmark dataset but uses real-world data collected during a 5 km experiment.
Key Insights and Ideas
- Doppler Measurements Aren't Immune: While often considered more robust than pseudo-range, Doppler measurements are also susceptible to NLOS and multipath effects, especially in challenging environments. The paper demonstrates that removing Doppler outliers can further improve localization.
- NFA for Outlier Detection is Powerful: The "a contrario" approach with NFA is effective because it's free from arbitrary threshold settings and robust to the choice of model parameters. It directly optimizes for the best subset of data.
- RBPF is Key for State Estimation: By splitting the state into linear and non-linear components, RBPF efficiently estimates both position and velocity, which is critical for accurate navigation, especially when dealing with noisy measurements.
- Combining Outlier Detection and RBPF is Synergistic: The paper shows that integrating the NFA-based outlier detection before the RBPF estimation step significantly boosts localization performance, outperforming methods that don't remove outliers or only remove pseudo-range outliers.
Key Results
- The proposed RBPF coupled with NFA outlier detection (RBPF + NFA (PR,Dp)) achieved the best localization performance.
- For errors less than 6m, RBPF + NFA (PR,Dp) achieved a 90.11% accuracy, significantly higher than the baseline EKF (71.66%) or PF (55.02%).
- For errors less than 3m, RBPF + NFA (PR,Dp) achieved 61.96% accuracy, outperforming other methods.
- The paper demonstrates that removing Doppler outliers, in addition to pseudo-range outliers, leads to better results, especially for filters that don't explicitly filter velocity (like PF).
- The NFA outlier detection method showed high accuracy (around 97%) and precision (around 97-98%) in classifying measurements.
Limitations and Future Work
- Computational Cost: While RBPF is more efficient than standard PF for high-dimensional states, the outlier detection step (Algorithm 1) involves iterative testing, which can be computationally intensive. The authors mention controlling this with the
Ntestparameter. - Parallelization: The outlier detection loop in Algorithm 1 can be parallelized, and the prediction step of RBPF can be processed concurrently with outlier detection, suggesting avenues for optimization.
- Stronger Coupling: The authors propose investigating a stronger coupling between the particle filtering and the a contrario estimation, potentially allowing the outlier detection to provide better input for the particle filter.
- Advanced Observation Models: Future work could involve using more sophisticated observation models that account for atmospheric effects, which might help in better understanding and potentially correcting detected outliers.
Summary updated