Abstract
Driving comfort is a crucial consideration in the automotive industry. In the realm of autonomous driving, comfort has always been a factor that requires continuous improvement. A common approach to improving driving comfort is through the optimization of local path planning. Nevertheless, it is imperative to recognize that macroscopic factors, including traffic flow and road conditions, wield a substantial influence on comfort. For instance, complex traffic scenarios increase the possibility of emergency braking, thereby affecting comfort. Consequently, investigating the intricate interplay between comfort and global path planning becomes essential. This paper introduces a methodology and framework for predicting driving comfort by leveraging road information. The study established a road information-driving comfort dataset and devised prediction models using multi-head attention mechanism. The ensuing discussion elucidates the practical application of the model in path planning through examples and tests. Following the path optimized by the model, the vehicles exhibited a reduction in jerk. This research predicted driving comfort based on road information and integrated it with global path planning, which holds significant implications for autonomous driving navigation systems and provides a valuable reference for related research.
Introduction
In the development of autonomous vehicles (AVs), "comfort" has traditionally been treated as a local optimization problem. Engineers typically employ Model Predictive Control (MPC) or trajectory smoothing to ensure gentle steering and braking once a path is determined. However, this approach overlooks a critical reality: no amount of smooth control can compensate for a poorly chosen route. A global path-planner that directs an AV through congested intersections or areas with high pedestrian activity inevitably leads to "jerk"—the rapid change in acceleration—due to frequent emergency braking and stop-and-go cycles. This research shifts the paradigm from local maneuvers to macroscopic route selection, treating comfort as a predictable outcome of road features like traffic density, speed limits, and intersection frequency.
Methodology
The proposed Autonomous Driving Comfort Prediction (ADCP) model utilizes a Multi-Head Attention (MHA) mechanism to capture the complex interplay between environmental variables. Unlike standard neural networks, MHA effectively models "feature crosses"—for instance, how a sharp turn's impact on comfort is amplified during heavy rain. The model's predictions are integrated into a modified A* search algorithm, enabling navigation systems to balance comfort against traditional metrics like travel time and distance.
Experimental Setup and Datasets
The study utilized a hybrid dataset combining real-world observations and high-fidelity simulations:
- Real-World Data: 274 datasets were collected across urban fast roads, highways, and rural routes in China using a test vehicle equipped with high-precision inertial measurement units (DS-IMU) and a SIRIUS XHS data acquisition system.
- Simulation Data: 252 datasets were generated via the 51sim-one platform, focusing on complex micro-scenarios such as traffic circles and busy intersections.
- Human Evaluation: 33 volunteers provided subjective comfort scores (HES) to ground objective sensor data (jerk and acceleration) in actual passenger experience.
Results and Discussion
The ADCP model demonstrated superior performance in predicting comfort levels across diverse scenarios. By prioritizing the "Most Comfortable" mode in global path planning, the system achieved significant improvements:
- Jerk Reduction: Longitudinal jerk was reduced by approximately 15%, while lateral jerk saw a 9% decrease.
- Passenger Satisfaction: Subjective human comfort scores improved by 13%.
- Model Accuracy: The attention-based model achieved a Mean Absolute Error (MAE) of 0.0908 for jerk, outperforming baseline BP neural networks and XGBoost models.
The results confirm that macroscopic environmental factors—such as traffic flow and road type—have a more substantial influence on comfort than previously recognized. Choosing a route with fewer intersections is often more effective than optimizing braking at those intersections.
Conclusion
This research introduces a framework for predicting driving comfort based on road information and integrating these predictions into global path planning. While current limitations include the difficulty of obtaining real-time road roughness data and the need for larger datasets, the study provides a valuable foundation for next-generation AV navigation systems. Future work will focus on dynamic weighting between travel time and comfort to cater to varying passenger preferences.
Summary updated