视觉融合下的复杂路况车辆换道决策模型

Vehicle Lane Change Decision Model for Complex Road Conditions with Visual Fusion

车辆换道是一个融合车辆行为及其周围环境的多因素动态过程,不合理的换道行为会招致严重的交通事故.现有安全防撞模型未考虑换道时多车辆间的耦合关系,且换道车辆的感知方式有一定局限性,在需强制性换道时决策效果不理想.因此本文提出视觉融合下的复杂路况车辆换道决策模型.该模型以单视角补偿双目视差计算的视觉融合方法感知参与换道的多车辆,解决了因车辆遮挡覆盖等非线性运动导致的多目标车辆轨迹预测误差问题,通过提取出三维的换道行为参数输入RBF神经网络进行可行性安全评估,可使车辆在合适时机进行换道.实验结果表明,提出的车辆换道决策模型在相同实验环境下相较单目检测算法YOLOv3准确性可提升5.1%,相较选取的基准双目检测算法准确性提升0.7%,且对换道进行安全评估的预测综合准确率达97.33%,能满足自动驾驶车辆强制性换道需求.

Vehicle lane changing is a multi-factor dynamic process integrating vehicle behavior and its surrounding environment,and unreasonable lane changing behavior can incur serious traffic accidents.The existing safety collision prevention model does not consider the coupling relationship between multiple vehicles when changing lanes,and the perception mode of lane changing vehicles has certain limitations,which makes the decision effect unsatisfactory when mandatory lane changing is required.Therefore,this paper proposes a lane changing decision model for complex road conditions with visual fusion.The model perceives multiple vehicles involved in lane-changing with the visual fusion method of monovision compensation binocular parallax calculation,solves the problem of multi-target vehicle trajectory prediction error caused by nonlinear motion such as vehicle occlusion coverage,and enables vehicles to make lane-changing at the right time by extracting the three-dimensional lane-changing behavior parameters into RBF neural network for feasibility safety assessment.The experimental results show that the proposed vehicle lane change decision model can improve 5.1%accuracy compared to the monocular detection algorithm YOLOv3 and 0.7%accuracy compared to the selected benchmark binocular detection algorithm in the same experimental environment,and the comprehensive prediction accuracy of safety assessment for lane change can achieve 97.33%accuracy,which can meet the demand for mandatory lane change of autonomous vehicles.