基于超距雷达数据的车辆跟驰行为分析及建模

Analysis and Modeling of High Speed Car-following Behavior Based on Ultra-range Radar Data

针对高速公路车辆跟驰特性及速度预测问题,基于超距雷达数据分析了高速公路车辆跟驰特性,建立起基于长短期记忆(long short term memory, LSTM)的跟驰速度预测模型。首先,根据数据特点建立了处理规则并筛选跟驰序列。其次,根据车身长度将跟驰车划为大型车与小型车,分析了车辆在跟驰过程中速度、车速差、车间距和车头时距等参数的分布及相对变化关系。然后,将前车速度、位置差、上一时刻车头时距作为模型输入,跟驰车速度作为模型输出,构建了基于LSTM的跟驰速度预测模型,模型预测精度达到99.75%。最后,以高速公路数据为例进行验证,传统机器学习支持向量回归(support vector regression, SVR)模型的预测性能低于深度学习模型,LSTM模型的R2较SVR模型提升了4.37%;作为LSTM的结构简化变体,在相同的结构参数下,门控循环单元(gated recurrent unit, GRU)模型的预测性能并未提升,但训练速度较LSTM模型提高了28.48%。深度学习LSTM、GRU模型能够更精准地预测高速公路的车辆跟驰速度。

To explore the car-following characteristics and speed prediction of freeway vehicles,the characteristics of high-speed car-following were analyzed based on the ultra-range millimeter wave radar data of freeway,and based on long short term memory(LSTM),a car-following speed prediction model was established.Firstly,the processing rules of radar data and the screening conditions of car-following state were established.Secondly,The distribution and relative variation of parameters such as speed,speed difference,spacing and time headway were analyzed by dividing the following vehicles into large or small.Then,the car-following speed prediction model based on LSTM was constructed and its prediction accuracy reaches 99.75%.Finally,taking highway data as validation,the performance of support vector regression(SVR)model was lower than that of deep learning model,and the R 2 of LSTM is 4.37%higher than that of SVR.As a simplified variant of LSTM,the prediction performance of gated recurrent unit(GRU)model is not improved under the same structural parameters,but the training speed is 28.48%higher than that of LSTM.Deep learning LSTM and GRU models can more accurately predict car-following speeds on highways.