基于BP神经网络的纵向避撞安全辅助算法

发布时间：2018-03-23 04:38

本文选题：主动安全　切入点：BP神经网络　出处：《西安交通大学学报》2017年07期 　论文类型：期刊论文

【摘要】：针对复杂人-车-路交通环境下汽车前向防碰撞预警系统(FCWS)在实际工程应用中存在虚警漏警、可接受性差等问题,为提高FCWS对驾驶群体行为差异的适应性,提出了驾驶员行为特性自适应学习的纵向避撞安全辅助算法,建立了基于BP神经网络的闭环驾驶跟驰习惯模型。该网络模型以跟车车距、前车减速度、前车速度、自车速度、环境亮度、路面附着系数、紧急制动次数、本次驾驶时间为输入,采用动量梯度下降自适应学习率方法对网络模型进行训练,进而预测出自车待制动减速度。引入聚类算法思想,对训练样本集进行特征聚类,进一步改善了BP网络的预测性能,设计了激进、谨慎、新手3类典型驾驶群体,通过制动深度、期望碰撞时间倒数、应急反应时间来表征驾驶群体性特征,在稳态跟车过程中对不同驾驶群体的行为特性进行学习,建立起非线性输入输出映射关系知识库,进而预测出相应群体的待制动行为,实现差异化预警,可有效降低虚警或预警不及时现象。仿真结果表明,改进后的BP网络减小了训练过程中陷入局部极小的可能性,提高了网络收敛速度以及预测精度,激进群体与新手群体待制动曲线下降的趋势相对更陡,激进群体跟车距离相对偏近,新手群体采取制动措施时所需相对车距相对偏远,以补偿较长的反应距离,从而验证了该理论模型对不同驾驶群体的适应性。
[Abstract]:In order to improve the adaptability of FCWS to the difference of driving group behavior, FCWS has some problems such as false alarm leakage, poor acceptability and so on, aiming at the problems of FCWSs in the complex human-vehicle-road traffic environment, in order to improve the adaptability of FCWSs to the difference of driving group behavior, so as to improve the adaptability of FCWSs. A longitudinal collision avoidance safety assistant algorithm based on adaptive learning of driver's behavior characteristics is proposed, and a closed loop driving car-following habit model based on BP neural network is established. The network model is based on the following distance, the deceleration of the front car, the speed of the front car, the self-speed of the vehicle. Environment brightness, road surface adhesion coefficient, emergency braking times, driving time are input, the momentum gradient descent adaptive learning rate method is used to train the network model. By introducing the idea of clustering algorithm, the characteristic clustering of training sample set is introduced, which further improves the prediction performance of BP network, and designs three typical driving groups, namely radical, cautious and novice driving groups. Through braking depth, countdown of expected collision time, emergency response time to characterize driving group characteristics, the behavior characteristics of different driving groups are studied in the process of steady-state vehicle following, and a knowledge base of nonlinear input-output mapping relationship is established. Then the behavior of the corresponding group is predicted and the differential early warning is realized. The simulation results show that the improved BP network reduces the possibility of falling into the local minima in the training process, which can effectively reduce the false alarm or the early warning untimely phenomenon, and the simulation results show that the improved BP network reduces the possibility of falling into the local minimum in the training process. The convergence speed and prediction accuracy of the network are improved. The trend of the decrease of the braking curve between the radical group and the novice group is steeper, the distance between the radical group and the vehicle is relatively close, and the relative distance between the vehicle and the vehicle is relatively remote when the novice group takes the braking measures. The adaptability of the model to different driving groups is verified by compensating the long reaction distance.
【作者单位】：中国科学院微电子研究所;中国科学院大学;中国科学院微电子研究所昆山分所;
【基金】：移动物联网关键技术与应用资助项目(XDA06040300)
【分类号】：TP183;U463.6

【参考文献】