The path inference filter: model-based low-latency map matching of probe vehicle data

Abstract: We consider the problem of reconstructing vehicle trajectories from sparse sequences of GPS points, for which the sampling interval is between 10 seconds and 2 minutes. We introduce a new class of algorithms, called altogether path inference filter (PIF), that maps GPS data in real time, for a variety of trade-offs and scenarios, and with a high throughput. Numerous prior approaches in map-matching can be shown to be special cases of the path inference filter presented in this article. We present an efficient procedure for automatically training the filter on new data, with or without ground truth observations. The framework is evaluated on a large San Francisco taxi dataset and is shown to improve upon the current state of the art. This filter also provides insights about driving patterns of drivers. The path inference filter has been deployed at an industrial scale inside the Mobile Millennium traffic information system, and is used to map fleets of data in San Francisco, Sacramento, Stockholm and Porto.

• The paper introduces a probabilistic path inference filter that integrates Bayesian CRF modeling with efficient graph search to accurately reconstruct sparse GPS trajectories.
• The paper demonstrates significant improvements in map matching accuracy and real-time performance, achieving notably low point and path misses at varying sampling rates.
• The paper validates its approach using Mobile Millennium project data, highlighting the algorithm’s scalability, robustness in urban settings, and practical implications for traffic estimation.

Overview of the Path Inference Filter for Map Matching

The paper explores a novel class of algorithms, termed the Path Inference Filter (PIF), tailored for the real-time mapping of GPS data. The primary focus is on efficiently reconstructing vehicle trajectories from sparse GPS observations, sampled at intervals ranging from 10 seconds to two minutes.

Core Contributions

The PIF stands out due to its probabilistic framework grounded in Bayesian principles, effectively combining observation and driver models into a Conditional Random Field (CRF). This integration allows the PIF to identify the most probable trajectories, enhancing accuracy even with low-frequency GPS data. The technique surpasses existing map-matching approaches, such as those based on deterministic, geometric, or simplistic probabilistic methods.

Algorithmic Design

The PIF's design is predicated on several key components:

1. Map Matching: GPS coordinates are projected onto candidate states within the road network, accounting for potential noise and errors.
2. Path Discovery: It computes feasible paths between states, leveraging efficient graph search algorithms, such as A*, ensuring adherence to vehicle dynamics.
3. Filtering: Through the use of a CRF, it assigns probabilistic weights to paths and states, smoothing trajectories by considering entire sequences rather than isolated observations.

Empirical Evaluation

The paper substantiates the efficacy of the PIF through rigorous evaluation using field data from the Mobile Millennium project. Key findings include:

• Accuracy: The PIF achieves notably low path and point misses at varying sampling rates, ranging from 1-second intervals to two-minute intervals.
• Robustness: The algorithm effectively handles urban environments, mitigating issues inherent in simpler methods.
• Scalability: A significant highlight is the algorithm's real-time capability, efficiently processing several hundred GPS observations per second.

Learning Architectures

Training the PIF involves both supervised and unsupervised methods. The supervised learning maximizes likelihood using known trajectories, while unsupervised learning leverages Expectation-Maximization, enabling training without explicit ground truth data.

Practical Implications and Future Developments

The deployment of PIF at scale in the Mobile Millennium system underscores its industrial relevance, providing insights into driving patterns and supporting traffic estimation. Future work might focus on integrating advanced driver behavior models or further optimizing computational performance.

Overall, the PIF marks a considerable advancement in map-matching techniques, offering a robust, probabilistic alternative capable of handling the intricacies of real-world GPS data and complex urban road networks.