摘要
自动驾驶车辆可以通过数据驱动模型较好地学习人类驾驶员的跟驰行为,但单纯的学习并不能发挥自动驾驶车辆反应更敏捷的特性。文中利用NGSIM数据集开发一种基于零反应时间数据的跟驰行为学习模型。首先,基于人类驾驶行为数据建立反应时间预测的神经网络模型,预测每条人类跟驰轨迹数据每个时间步的反应时间,并在原轨迹中剪除反应时间内的数据,进而重构样本数据,获得近似于零反应时间、更符合自动驾驶车辆特性的样本集。在此基础上采用LSTM架构,建立基于新学习样本的跟驰行为模型(LSTM-0RT)。仿真对比发现:LSTM-0RT跟驰模型比传统LSTM模型提前50 s收敛,且速度变化趋势与前车基本一致,充分体现反应速度快的特点;在混驶环境测试中,采用LSTM-0RT模型的自动驾驶车辆比例越大,跟驰车队的渐进稳定性越高,车流波动的影响范围越小;交通流特性分析显示LSTM-0RT模型在不同交通流密度下的适用性明显优于LSTM模型;车头时距指标测算也表明LSTM-0RT模型具有更高的跟驰安全性。
Although autonomous driving vehicles can better learn the car-following behavior of human drivers through data-driven models,learning alone cannot give play to the more agile characteristics of autonomous driving vehicles.This paper uses the NGSIM data set to develop a car-following behavior learning model based on zero response time data.Firstly,a neural network prediction model for reaction time is established based on human driving behavior data,to predict the reaction time of each human car-following trajectory data at each time step,cut the response time data from the original trajectory,and then reconstruct the sample data to obtain a sample set that is more in line with the characteristics of autonomous vehicles with zero response time.On this basis,the LSTM architecture is adopted to establish a car-following behavior model(LSTM-0RT)based on new learning samples.Through simulation comparison,it is found that the LSTM-0RT car-following model converges 50s earlier than the LSTM car-following model,and the speed change trend is basically the same as that of the preceding vehicle,which fully reflects the characteristics of fast response.In the test of the mixed driving environment,the larger the proportion of autonomous vehicles using the LSTM-0RT model,the higher the gradual stability of the car-following fleet and the smaller the influence range of traffic fluctuations.Analysis of traffic flow characteristics shows that the applicability of the LSTM-0RT model under different traffic flow densities is significantly better than that of the LSTM model.The calculation of headway indicators also shows that the LSTM-0RT model has higher car-following security.
作者
程陆
柏海舰
CHENG Lu;BAI Haijian(School of Automotive and Transportation Engineering,Hefei University of Technology,Hefei 230009,China)
出处
《交通科技与经济》
2021年第6期7-16,共10页
Technology & Economy in Areas of Communications
基金
国家自然科学基金面上项目(52072108)。
关键词
混驶环境
自动驾驶
跟驰模型
LSTM
反应时间
mixed driving environment
autonomous driving
car-following model
LSTM
reaction time
