An autonomous driving decision-making framework for joint prediction and planning in unsignalized intersection scenarios

Abstract

Unsignalized intersections present one of the most challenging environments in autonomous driving due to their complex traffic scenarios. Safely and efficiently navigating these uncertain settings remains a significant research hurdle. To tackle this issue, this paper proposes an End-to-End Autonomous Driving Decision Framework (EJPP) based on the interactive fusion of prediction and planning modules. The framework accurately predicts future trajectories of surrounding vehicles to facilitate optimal path planning. The EJPP framework consists of prediction and planning modules. The prediction module integrates vehicle acceleration as implicit behavioral intent and utilizes Pearson correlation coefficients to comprehensively consider complete information interactions among vehicles, thereby mitigating potential trajectory uncertainties. Moreover, a temporal attention mechanism is incorporated to capture critical temporal features from historical trajectories and enhance prediction accuracy. Within the planning module, trajectories of autonomous vehicles are planned separately in the lateral and longitudinal directions using the Frenet coordinate system. By integrating vehicle dynamics into cost functions encompassing safety, comfort, and efficiency, both soft and hard constraints are designed to optimize the optimal route. The proposed framework is validated through closed-loop training across various traffic flow scenarios to assess its performance. Results indicate that the framework enables autonomous vehicles to traverse unsignalized intersections safely and efficiently.