TL;DR
This paper introduces a novel data fusion method for driverless cars that combines GPS, IMU, and Dead Reckoning (DR) data. By using a set of predictive models and occupancy grid constraints, the method significantly improves localization accuracy and robustness, outperforming existing state-of-the-art approaches.
What Problem Does This Paper Address?
Accurate and reliable localization is a cornerstone for autonomous navigation in driverless cars. While GPS offers absolute positioning, it's susceptible to environmental factors like signal occlusion and multipath interference, leading to inaccuracies. Dead Reckoning (DR), which uses vehicle dynamics, can provide continuous positioning but suffers from accumulated errors over time. Integrating these sensors (GPS, IMU, DR) is crucial, but existing fusion methods like Extended Kalman Filters (EKF) and Unscented Kalman Filters (UKF) struggle with uncertainties in noise characteristics and can be computationally intensive. This paper aims to address the challenge of achieving robust, high-precision localization in complex urban environments by developing a more effective data fusion technique.
Methods
The core of this paper is a novel data fusion method that leverages a set of predictive models and occupancy grid constraints. Here's a breakdown:
- ARMA Predictive Models: Instead of relying on a single model order, the authors propose using a set of Autoregressive Moving Average (ARMA) models with varying structural orders (1st, 2nd, and 3rd order in this case) to predict future positions from both GPS-IMU and DR data. This approach avoids the difficult task of pre-determining the optimal ARMA model order.
- Occupancy Grid Constraints & Spatial Consensus Check: Raw predictions and current measurements are projected onto an occupancy grid map. Outliers are then removed through two mechanisms:
- Grid Constraints: Points falling into grid cells with a majority of other points (inliers) are retained, while sparse points are classified as outliers. The grid size (0.2m x 0.2m) is empirically determined to be optimal.
- Spatial Consensus Check: An iterative filtering process refines the inliers within grid cells to identify the most accurate centroid.
- Data Fusion: After outlier removal, the remaining inlier data points are fused to produce the final, accurate localization estimate. The standard deviation of the fused data can be controlled by the grid size.
Datasets
The paper primarily uses data collected from on-site experiments conducted with their driverless car in real urban scenarios. They also reference datasets from a previous study ([15]) for comparative analysis with existing methods.
Key Insights and Ideas
- Order-Agnostic Prediction: Using multiple ARMA models of different orders simultaneously bypasses the complex and often suboptimal process of selecting a single "best" model order. This makes the prediction step more robust.
- Grid-Based Outlier Rejection: The occupancy grid acts as a powerful spatial filter, effectively identifying and discarding erroneous measurements or predictions that deviate significantly from the consensus. This is a key innovation for improving data quality before fusion.
- Controlled Uncertainty: The ability to control the standard deviation of the fused data by adjusting the grid size offers a practical way to tune the localization precision based on environmental requirements.
- Smoothness in Challenging Conditions: The method demonstrates remarkable resilience to GPS signal interruptions and multipath effects, producing smooth trajectories even when raw GPS-IMU data exhibits significant jumps.
Key Results
- Improved Accuracy: The proposed method significantly reduces accumulated position errors from DR and smooths out biases from GPS-IMU, achieving a final average position error of 0.18 m. This is notably better than stand-alone GPS-IMU (0.47 m) and DR (0.49 m) and also outperforms a state-of-the-art NLM (Nonlinear Model) filter, which had an RMSE of 0.8865 m for position.
- Reduced Computational Load: The method exhibits superior performance in terms of computational efficiency, with a cycle duration of 0.005 ms, significantly faster than the NLM + CF (Complementary Filter) approach (1.0 ms).
- Robustness: Experimental results show that the fusion method maintains smooth and accurate trajectories even during periods of GPS signal occlusion or multipath interference.
Limitations and Future Work
- Degradation with Prolonged Bias: The authors acknowledge that if GPS-IMU data exhibits discontinuity jumps for an extended period, the fusion accuracy might degrade. They suggest that in such scenarios, a policy to temporarily block the GPS-IMU data source might be necessary.
- Turning-About Maneuvers: While generally effective, the method still needs improvement for complex maneuvers like turning about, possibly due to azimuth errors from DR when crossing intersections. This suggests future work could focus on enhancing DR's performance in these specific situations.
- Further Improvement: The authors indicate that their method "still needs to be further improved" for certain scenarios, hinting at ongoing research and development.
Summary updated