Aggressive Driving Saves More Time? Multi-task Learning for Customized Travel Time Estimation

Ruipeng Gao, Xiaoyu Guo, Fuyong Sun, Lin Dai, Jiayan Zhu, Chenxi Hu, Haibo LiIJCAI 2019 · DOI

TL;DR

This paper introduces Customized Travel Time Estimation (CTTE), a novel multi-task learning framework that fuses GPS, smartphone inertial data, and road network information to predict both general traffic speed and personalized travel times. By accounting for individual driving behaviors, CTTE aims to provide more accurate travel time estimates than existing methods.

What Problem Does This Paper Address?

Estimating travel time is crucial for many location-based services, from ride-hailing to route planning. While existing methods have improved accuracy, they often rely on coarse GPS data and struggle to capture personalized driving behaviors. Drivers can exhibit "aggressive" driving (e.g., speeding, sudden braking, frequent lane changes) which can significantly impact travel time, yet this aspect is largely ignored by current models. This paper addresses the challenge of accurately estimating travel time by considering both general traffic conditions and the unique driving patterns of individual users, aiming to answer the question: "Does aggressive driving actually save time?"

Methods

CTTE employs a multi-task learning approach with several key components:

  • Data Fusion: It integrates three primary data sources:
  • GPS Trajectories: Provide location and time information.
  • Smartphone Inertial Data (IMU): Accelerometer and gyroscope data are used to infer fine-grained driving events.
  • Road Network: Includes topological information, speed statistics, and query distribution.
  • Pose Estimation: A Principal Component Analysis (PCA) algorithm is used to estimate the smartphone's pose within the vehicle's coordinate system, crucial for accurately interpreting inertial data.
  • Aggressive Driving Detection: Inertial data, after pose correction, is processed to identify three specific aggressive driving events: speeding, sudden brake, and frequent lane shifts.
  • Link Traffic Database: A database is constructed containing link topology, speed statistics, and query amounts for each road segment.
  • Multi-task Learning Architecture: A deep recurrent neural network (RNN) is used. It first predicts general traffic speed on each road link. Simultaneously, it learns a driver-specific scale factor based on detected aggressive driving behaviors. This scale factor is then used to adjust the predicted travel time, personalizing it for each driver. The model is trained to minimize a combined loss for both speed prediction and travel time estimation.

Datasets

The paper utilizes two large-scale, real-world traffic datasets collected via crowdsourcing from the DiDi platform, covering the cities of Beijing and Shanghai. These datasets include:

  • Vehicle GPS trajectories
  • Smartphone inertial data
  • Road network information
  • Auxiliary information (weather, holiday, day of the week)

The GPS points in each trace were projected onto road links using a map-matching algorithm.

Key Insights and Ideas

  • Personalized Driving Matters: The core insight is that individual driving styles, particularly aggressive behaviors, significantly influence travel time and cannot be ignored for accurate estimation.
  • Inertial Data is Key for Fine-Grained Events: While GPS is useful for general location, smartphone inertial sensors are essential for detecting subtle, rapid driving maneuvers like lane changes and sudden brakes.
  • Road Network Topology is Crucial: Incorporating road network topology improves traffic speed prediction accuracy, suggesting that understanding the structure of the road network is more effective than simply using link IDs.
  • Multi-task Learning for Efficiency: Predicting both general traffic speed and personalized travel time within a single framework is an efficient way to leverage shared information and improve overall performance.
  • Driver-Specific Scale Factor: Using a learned scale factor based on driving behavior is a more nuanced approach than simply using a driver ID, allowing for better generalization to new users and capturing similar behaviors across different drivers.

Key Results

  • Improved Travel Time Estimation: CTTE achieved the lowest Mean Absolute Percentage Error (MAPE) among the compared methods on both Beijing and Shanghai datasets, indicating superior accuracy in travel time prediction.
  • Link Speed Prediction: Incorporating road network information reduced the Mean Squared Error (MSE) for link speed prediction from 22.5 to 16.9. The attention mechanism further improved this to 16.2.
  • Aggressive Driving Detection: The system achieved high precision and recall for detecting aggressive driving events, with all metrics above 86%.
  • Comparison with Baselines: CTTE outperformed SVR and DeepTTE in terms of MAPE, while also maintaining competitive MSE and MAE. For instance, on the Beijing dataset, CTTE achieved a MAPE of 22.6% compared to SVR's 83.1% and DeepTTE's 25.8%.

Limitations and Future Work

  • Data Sparsity: The paper notes that IMU data can be sparse on current platforms, limiting its impact. Future work could explore more fine-grained IMU data for greater improvements.
  • Generalization to New Users: While the driver-specific scale factor is an improvement, further research could explore even more robust methods for handling new users with minimal historical data.
  • Other Aggressive Events: The current work focuses on three specific aggressive driving events. Future work could expand to identify other pivotal events, such as not maintaining a safe distance from the preceding vehicle, potentially using computer vision from dashboard cameras.
  • Hyperparameter Robustness: Experiments showed robustness to dropout values and hidden state sizes, but further exploration of other hyperparameters could be beneficial.

Summary updated

BibTeX
@inproceedings{gao2019aggressive,
  author = {Ruipeng Gao and Xiaoyu Guo and Fuyong Sun and Lin Dai and Jiayan Zhu and Chenxi Hu and Haibo Li},
  title = {{Aggressive Driving Saves More Time? Multi-task Learning for Customized Travel Time Estimation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/234},
}
Figure