关键词: 自动驾驶, 安全场, 轨迹规划, 轨迹预测

Abstract:

A local trajectory planning method based on the driving safety field was proposed to address the challenge of trajectory planning for autonomous vehicles during lane-changing maneuvers caused by the uncertain motion states of surrounding dynamic obstacles. The method adopted a decoupled lateral-longitudinal planning framework. In lateral planning, a family of candidate paths was generated using B-spline curves, and a multi-objective evaluation function was formulated by incorporating virtual risk forces, path length, and curvature. An adaptive weighting strategy was applied to select the optimal path, thereby enhancing driving safety and responsiveness to sudden scenarios. In longitudinal planning, a long short-term memory network was first employed to predict the trajectories of surrounding vehicles, and a safety region was constructed that accounted for prediction uncertainty. A rough speed profile was then obtained via dynamic programming within a spatiotemporal graph, followed by quadratic programming using quintic polynomials to generate a smooth and feasible velocity trajectory. The results show that, relative to the Gaussian potential field baseline, the proposed method yields a 22.45% reduction in lane-change risk in relatively safe scenarios and a 30.98% reduction in relatively hazardous scenarios.

Key words: autonomous vehicles, safety field, trajectory planning, trajectory prediction