基于GCN和QP的智能车辆换道决策规划

Lane Change Decision Making and Planning of Intelligent Vehicles Based on GCN and QP

考虑动态驾驶场景下车辆间的交互影响,提出了一种基于图卷积网络和二次规划的智能车辆自主换道行为决策与运动规划方法.首先将感兴趣区域进行分层建模,以图结构数据的形式对驾驶场景的全局和局部动态交互信息进行聚合,通过图卷积网络输出自车应采取的驾驶行为决策指令,然后与运动规划模块结合,基于局部子图划分可通行区域,构建并求解二次规划模型,得到满足运动学约束的无碰撞运动轨迹,最终完成无碰撞自主换道.对提出的方法进行了仿真实验与实车验证,实验结果证明该方法的性能要优于传统的规划方法,具有更好的实验成功率以及场景泛化性能.

To solve the problem of the interaction effects between vehicles in dynamic driving scenarios,an autonomous lane change decision making and motion planning method were proposed for intelligent vehicles based on graph convolution network(GCN)and quadratic programming(QP).Firstly,some interested regions were hierarchically modeled,and the global and local dynamic interaction information of the driving scene were aggregated in a form of graph-structured data,and the driving behavior instructions should take in the ego vehicle were output with the GCN.Then,combined with motion planning module,the free spaces were divided based on the local sub-graph,a quadratic programming model was constructed and solved to obtain collision-free motion trajectory satisfied with kinematics constraints,completing the autonomous lane change without colli-sion finally.The results of simulation experiments and real vehicle verification show that the proposed method can provide better performance than the conventional decision making and motion planning method,showing better experimental success rate and scene generalization performance.