Lane Departure Prediction Based on Closed-Loop Vehicle Dynamics (2112.10379v3)

Abstract: An automated driving system should have the ability to supervise its own performance and to request human driver to take over when necessary. In the lane keeping scenario, the prediction of vehicle future trajectory is the key to realize safe and trustworthy driving automation. Previous studies on vehicle trajectory prediction mainly fall into two categories, i.e. physics-based and manoeuvre-based methods. Using a physics-based methodology, this paper proposes a lane departure prediction algorithm based on closed-loop vehicle dynamics model. We use extended Kalman filter to estimate the current vehicle states based on sensing module outputs. Then a Kalman Predictor with actual lane keeping control law is used to predict steering actions and vehicle states in the future. A lane departure assessment module evaluates the probabilistic distribution of vehicle corner positions and decides whether to initiate a human takeover request. The prediction algorithm is capable to describe the stochastic characteristics of future vehicle pose, which is preliminarily proved in simulated tests. Finally, the on-road tests at speeds of 15 to 50 km/h further show that the pro-posed method can accurately predict vehicle future trajectory. It may work as a promising solution to lane departure risk assessment for automated lane keeping functions.

• The paper introduces a novel closed-loop vehicle dynamics model that improves lane departure risk prediction compared to traditional methods.
• It employs a physics-based approach with techniques like EKF and Kalman Predictor with Control to simulate vehicle states and steering dynamics.
• Simulations and on-road tests demonstrate the method’s superior accuracy over conventional CTRV models, bolstering automated driving reliability.

Overview of Lane Departure Prediction Using Closed-Loop Vehicle Dynamics

The paper "Lane Departure Prediction Based on Closed-Loop Vehicle Dynamics" presents a sophisticated approach to vehicle trajectory prediction crucial for autonomous driving systems, particularly in lane keeping scenarios. The research proposes an innovative application of a closed-loop vehicle dynamics model for predicting lane departure risks, emphasizing the essentiality of predicting vehicle trajectories accurately to enhance the safety and reliability of automated driving functions.

Summary of Methodology

The methodology presented leverages a physics-based approach, diverging from the traditional trajectory prediction methods by incorporating closed-loop control dynamics. The core process is divided into three phases: vehicle state estimation, motion prediction, and lane departure assessment. Vehicle state estimation employs an Extended Kalman Filter (EKF) to approximate current vehicle states derived from sensor outputs. Subsequently, a Kalman Predictor with Control (KPC) simulates future steering actions and vehicle states, furthering prediction accuracy by integrating actual lane keeping control laws. The Lane Departure Assessment (LDA) module evaluates probabilistic distributions of potential vehicle positions, ultimately determining the necessity of human intervention to avert lane departure.

Notable Results

The paper substantiates the validity of the proposed algorithm through both simulation and on-road testing. Numerical simulations demonstrate that the KPC algorithm can effectively predict vehicle trajectories, maintaining accuracy where conventional Constant Turn Rate and Velocity (CTRV) models fall short, particularly in situations where predictive assumptions can diverge, such as control input variability. On-road test trials indicate that the methodology remains effective over a speed range of 15 to 50 km/h, generally predicting accurate trajectories while outperforming CTRV in terms of providing a more precise risk assessment of potential lane departures.

Implications and Future Directions

The proposed method represents a significant step towards improving the risk assessment capabilities of automated driving systems. By integrating closed-loop dynamics, the solution addresses a critical need for accurate prediction over the long-term while retaining the practical efficiency of conventional physics-based approaches. Among the practical implications of this paper, it suggests that the inclusion of trajectory prediction grounded in vehicle control dynamics can attenuate unnecessary intervention and improve user acceptance of automated driving technologies.

Looking forward, this paper outlines potential trajectories for further exploration. Enhancements could involve incorporating considerations for unknown disturbances or actuator errors, which were only minimally addressed in the present paper. Furthermore, improving predictive models to include longitudinal vehicle dynamics and adaptive learning of control laws may present opportunities to boost the robustness and applicability of the proposed model. Integrating machine learning algorithms to complement the physics-based approach could also enhance the accuracy and adaptability of the algorithm in dynamic, real-world driving environments.

In conclusion, this research contributes a methodologically rigorous framework for lane departure prediction in automated driving systems. Its focus on leveraging existing control dynamics marks a meaningful progression within the domain of predictive automotive technology, offering a path forward for more reliable and trustworthy autonomous vehicle operation systems.