加权指数损失下长短时记忆网络换道意图识别模型

针对车道变换意图识别中数据源单一,传统序列模型难以捕获长序列范围内换道意图且存在长期依赖问题,提出一种结合时间信息加权指数损失函数的长短时记忆(long short-term memory,LSTM)车辆换道意图识别模型.首先,利用驾驶模拟舱、眼动仪进行高速公路驾驶实验,采集车辆运行数据和驾驶员眼动数据;然后,基于LSTM结构单元构建高速公路环境下车辆换道意图识别模型,提出基于时间信息加权的指数损失函数对模型权重进行优化;最后,利用车辆运行数据和驾驶员眼动数据对所提模型加以验证并与其他模型进行对比,所提模型换道识别的准确率为91.33％,宏平均精确率为89.04％,宏平均召回率为92.84％,宏平均F1值为90.33％.结果表明,长短时记忆网络对于长序列换道意图识别过程具有较好的分辨能力,提出的损失函数对模型权重优化具有良好的效果.

A LSTM Intent Recognition Model for Lane Change Based on Time-weighted Exponential Loss

Aiming at the single data source in lane change intention recognition, the traditional sequence model is difficult to capture the lane change intention in the long sequence range and there is a long-term dependency problem. A LSTM (long short-term memory) vehicle lane change model combined with time information weighted exponential loss function is proposed. First, the driving simulator and Eye tracker are selected to carry out the highway driving experiment. The driving data and the driver''s eye movement data are collected respectively. Second, based on the LSTM structural unit, the intention recognition model of lane changing under the highway environment is constructed, and the multi-exponential loss function based on time information weighting is proposed to optimize the weights. Finally, the proposed model is verified by vehicle driving data, eye movement data and compared with other models. The accuracy of the proposed lane-changing recognition is 96.78%. The accuracy rate is 95.72%. The recall rate is 95.83%. The F1 value is 95.73%. The results show that the long short-term memory network has a good capacity for the long-sequence lane change intent recognition process, and the proposed loss function achieves a positive effect on model weight optimization.