This paper proposes a Dynamic Bayesian Network framework for autonomous vehicles that enhances safety in cut-in maneuvers by integrating lateral evidence and probabilistic safety assessments, achieving superior crash avoidance in high-speed scenarios (9.22% crash rate) compared to baseline models in the JRC-FSM simulator.
Safety, Robustness, Prediction, Planning, Robotics
Kranthi Kumar Talluri, Anders L. Madsen, Galia Weidl
Aschaffenburg University of Applied Sciences, HUGIN EXPERT A/S, Aalborg University
Generated by grok-3
Background Problem
The increasing presence of autonomous vehicles (AVs) in real-world traffic necessitates robust safety models to handle dynamic and unpredictable conditions, particularly cut-in maneuvers that often lead to abrupt braking and collisions. Traditional heuristic and statistical models fail to adapt to real-world uncertainties, while advanced AI/ML approaches like deep learning and reinforcement learning suffer from high computational demands, lack of explainability, and reliance on perfect perception. This paper addresses the critical problem of safe lane change prediction and cut-in maneuver planning by leveraging probabilistic models to enhance safety and decision-making in automated driving systems.
Method
The proposed method utilizes a Dynamic Bayesian Network (DBN) framework to model and predict lane change maneuvers for safe cut-in scenarios in autonomous vehicles. The core idea is to integrate lateral evidence with safety assessments using three key probabilistic hypotheses: Lateral Evidence (LE), Lateral Safety (Safe_lat), and Longitudinal Safety (Safe). The DBN processes real-time data such as vehicle positions, lateral velocities, relative distances, and Time-to-Collision (TTC) to infer safety-critical events and produce safety decisions at each time step (updated at 10 Hz). The implementation involves modeling these hypotheses in the HUGIN software and integrating them into the JRC-FSM simulator to enhance the baseline CC human driver model. Key steps include: (1) LE computes lane change probabilities using a sigmoid function based on lateral velocity and distance; (2) Safe_lat evaluates lateral safety margins by checking clearance between vehicles; and (3) Safe assesses longitudinal safety using TTC and relative speed, triggering deceleration if unsafe conditions are predicted. The method focuses on early detection and proactive safety checks without modifying the underlying vehicle control logic.
Experiment
The experiments were conducted using the JRC-FSM simulator, a Python-based framework designed to evaluate safety algorithms under complex cut-in scenarios, with comparisons against baseline models like CC human driver, FSM, RSS, and Reg157. The setup included various low-speed (10-60 km/h) and high-speed (70-130 km/h) scenarios, testing parameters such as initial distance, lateral velocity, and relative speed between ego and cut-in vehicles. Metrics like crash avoidance rate, minimum TTC, deceleration, and jerk were used to assess performance. Results showed that the DBN model significantly outperformed others in high-speed scenarios, achieving a crash percentage of 9.22% compared to 25.30% for CC human driver, with early detection (1.5-3 seconds earlier) enabling smoother braking and reduced jerk. In low-speed scenarios, DBN’s performance was competitive (6.81% crashes) but not superior to FSM or RSS. The experimental design focused on critical scenarios to stress-test models, which is reasonable for safety validation, though it may overemphasize high-speed improvements. While the results align with the expectation of enhanced safety through probabilistic modeling, the lack of real-world testing and potential computational overhead at 10 Hz updates remain unaddressed limitations.
Further Thoughts
The early detection capability of the DBN model, which predicts lane changes 1.5-3 seconds earlier than baselines, is a significant advancement that could inspire further research into probabilistic models for other safety-critical tasks in autonomous systems, such as pedestrian detection or intersection navigation. However, a deeper exploration is needed on how this framework scales to real-world environments where sensor noise, diverse traffic behaviors, and computational constraints are more pronounced. An interesting connection could be drawn to the field of Human-AI Interaction, where the explainability of DBNs (as highlighted by the authors) could be leveraged to build trust in AV systems by providing transparent decision-making processes to regulators and end-users. Additionally, comparing DBNs with emerging foundation models for vision and decision-making in AVs could reveal whether probabilistic approaches remain competitive against data-driven, black-box models in terms of safety and efficiency. Future work could also explore hybrid models combining DBNs with reinforcement learning to balance interpretability and adaptability in dynamic traffic scenarios.