Integrated Decision and Motion Planning for Highways with Multiple Objects Using a Naturalistic Driving Study

Feng Gao, Zheng Xu, Qiuxia Hu, Hongwei LiuSensors 2024 · DOI

TL;DR

This paper introduces a novel two-layered decision and motion planning system for automated vehicles (AVs) on highways with multiple surrounding objects. By modeling a "cognitive safety area" derived from naturalistic driving data and employing an artificial potential field approach, the system effectively decouples environmental interactions from ego-vehicle dynamics, leading to safer, more efficient, and comfortable driving.

What Problem Does This Paper Address?

As automated vehicles become more intelligent, integrating various driving modules (perception, decision-making, planning) becomes crucial for enhanced performance and adaptability. However, managing the complex interactions between an AV and multiple surrounding vehicles in dynamic highway scenarios presents a significant challenge. Traditional rule-based systems struggle with the sheer number of possible situations, while purely optimization-based or end-to-end learning approaches can be computationally expensive or lack interpretability. This paper aims to bridge this gap by proposing an integrated system that can handle complex multi-object highway driving scenarios effectively and efficiently.

Methods

The core of the proposed solution is a two-layered decision and motion planning strategy.

  • Outer Layer (Cognitive Safety Area): This layer focuses on the interaction between the AV and its surrounding environment. A "cognitive safety area" is defined around the ego vehicle (EV). The size and dynamics of this area are determined by analyzing naturalistic driving behavior, specifically focusing on human driver attention during lane changes. An artificial potential field (APF) method is used to model the influence of all surrounding objects on this safety area.
  • Inner Layer (Ego Vehicle Dynamics): This layer handles the dynamic control of the EV within the safety area. The interaction between the EV and the safety area is simplified into a spring-damping system. This allows for analytical calculation of the desired dynamic states (acceleration) of the EV, improving computational efficiency.

Key aspects of the method include:

  • Decoupling: The two-layered structure separates the complex environmental interactions from the direct control of the EV.
  • Naturalistic Driving Data: The parameters for the cognitive safety area are derived from real-world driving data, aiming to mimic human driving behavior.
  • Artificial Potential Field (APF): Used to represent the forces exerted by surrounding vehicles on the safety area, allowing for a unified approach to multi-object interactions.
  • Spring-Damping System: Simplifies the EV's motion planning by treating its interaction with the safety area as a mechanical system, enabling analytical solutions.
  • Lateral Motion Decision Logic: A specific logic is designed for the safety area's lateral movement during lane changes, based on projected forces from surrounding vehicles and lane social forces.

Datasets

  • HighD Dataset: This dataset, collected by drones, provides pre-processed trajectory data for vehicles on German highways. It was used for parameter estimation and validation of the proposed method, particularly for lane change scenarios. The paper notes that the selected data covers a wide range of lane change scenarios.

Key Insights and Ideas

  • Mimicking Human Driving: By deriving the "cognitive safety area" from naturalistic driving data, the system aims to replicate human-like driving behavior, potentially leading to more comfortable and predictable AVs.
  • Simplification through Abstraction: The concept of a "cognitive safety area" acts as an abstraction layer, simplifying the complex interactions with multiple surrounding vehicles into a more manageable problem.
  • Analytical Efficiency: Modeling the EV's interaction with the safety area as a spring-damping system allows for analytical solutions, significantly improving computational efficiency compared to purely numerical optimization methods.
  • Customizable Driving Style: The conversion factor 'a' in the APF model is identified as a parameter that can be adjusted to customize the driving style, allowing for more aggressive or conservative behavior.

Key Results

The paper presents validation through both open-loop and closed-loop tests, comparing the proposed Integrated Driving Decision and Motion Planning (IDP) method with the Intelligent Driver Model (IDM) and Nonlinear Model Predictive Control (NMPC).

  • Open-Loop Tests (Naturalistic Driving Data):
    • The proposed method showed better fitting accuracy compared to IDM, indicating that the EV controlled by IDP more closely resembles a human driver.
  • IDP demonstrated better accuracy in predicting the lane change moment delay compared to IDM.
  • Closed-Loop Tests (Simulation):
    • Safety: IDP achieved a better overall safety index than IDM and NMPC.
  • Efficiency: IDP showed a slightly better average efficiency index than IDM.
  • Comfort: IDP's comfort performance was comparable to or slightly better than IDM, and significantly better than NMPC, which was found to be more conservative and uncomfortable.
  • Computational Cost: NMPC required significantly more computation resources (over seven times that of IDP and IDM) and had a longer average period.
  • Multi-Task Dynamic Scenario:
    • All three methods could pass the complex scenario without collision.
  • IDP was found to be faster and drove in a more stable manner, implying a more comfortable experience.

Limitations and Future Work

  • Ethical and Legal Requirements: The paper acknowledges that ethical and legal considerations, particularly for unavoidable accident scenarios, need to be further studied and integrated into the system design.
  • Dataset Diversity and Robustness: While the HighD dataset is valuable, the authors suggest testing the proposed method with more diverse datasets and considering additional factors like weather to demonstrate its robustness and generalization capabilities.
  • Real-World Validation: The current validation is primarily through simulation and open-loop tests on naturalistic data. Further real-world testing would be crucial to confirm its performance in actual driving conditions.

Summary updated

Abstract

With the rise in the intelligence levels of automated vehicles, increasing numbers of modules of automated driving systems are being combined to achieve better performance and adaptability by reducing information loss. In this study, an integrated decision and motion planning system is designed for multi-object highways. A two-layer structure is presented to decouple the influence of the traffic environment and the dynamic control of ego vehicles using the cognitive safety area, the size of which is determined by naturalistic driving behavior. The artificial potential field method is used to comprehensively describe the influence of all external objects on the cognitive safety area, the lateral motion dynamics of which are determined by the attention mechanism of the human driver during lane changes. Then, the interaction between the designed cognitive safety area and the ego vehicle can be simplified into a spring-damping system, and the desired dynamic states of the ego vehicle can be obtained analytically for better computational efficiency. The effectiveness of this on improving traffic efficiency, driving comfort, safety, and real-time performance was validated using several comparative tests utilizing complicated scenarios with multiple vehicles.
BibTeX
@inproceedings{gao2024integrated,
  author = {Feng Gao and Zheng Xu and Qiuxia Hu and Hongwei Liu},
  title = {{Integrated Decision and Motion Planning for Highways with Multiple Objects Using a Naturalistic Driving Study}},
  booktitle = {Sensors},
  year = {2024},
  doi = {10.3390/s25010026},
}
Figure