Trajectory Planning for Highway Autonomous Driving by Integrating a Risk-Pulse Driving Risk Field with a Spatiotemporal Semantic Corridor

This study proposes a highway trajectory-planning method integrating a risk-pulse driving risk field (RPDRF) with a spatiotemporal semantic corridor (SSC) to strengthen forward-looking risk quantification and its transfer between behavioral decisions and trajectory optimization. Trajectory predictions inform dynamic, static, unified, and risk evolution fields. Lane-level risk aggregation supports multipolicy maneuver selection, while bounded field-gradient targets and a convex quadratic risk cost guide piecewise Bézier trajectories within the SSC. Evaluation uses 27 field-data-derived highway segments covering car-following, lane-changing, overtaking, and curved-road driving. Mean planning-stage time is 167.82 ms, and the overtaking composite score is 0.918. The four-scenario macro-average score, minimum scenario score, and mean normalized metric score are 0.722, 0.533, and 0.745, respectively, exceeding the baseline summaries. Generalization evaluations report the highest composite-score point estimates in lane changing and overtaking for both SQM-W-2 and US101 and in curved-road driving for US101. The complete configuration has the highest ablation macro-average score; the lower bounds of the 95% descriptive intervals for the main macro-average score differences and the minimum differences in leave-one-segment-out checks are all positive. These results indicate potential for adaptation across datasets and consistent comparative performance under the evaluated reconstructed highway conditions. The framework connects spatiotemporal risk perception and semantic constraints, providing an interpretable approach to risk-guided trajectory planning in complex highway scenarios.