人机混驾环境下基于LSTM的无人驾驶车辆换道行为模型

道路系统中的人机混驾交通环境是指人工驾驶车辆与自动驾驶车辆混合运行的交通环境，其中换道行为建模是人机混驾环境下无人驾驶车辆行为研究的热点。基于深度学习理论，构建人机混驾环境下基于长短期记忆神经网络的无人驾驶车辆换道行为模型（Long-short-term-memory-based Autonomous Vehicles Lane Changing，LSTM-LC）。通过研究人工驾驶车辆在换道过程中与周边车辆的相互作用，对换道行为影响因素进行分析；同时，为了提升模型的迁移性，引入道路横向偏移量信息。结合LSTM神经网络的输入要求，使用美国公开交通数据集Next Generation SIMulation（NGSIM）构建换道行为样本库。针对LSTM-LC模型，以均方差MSE作为损失函数，使用RMSprop优化方法进行训练，对LSTM网络结构、历史序列长度N及训练样本量3个重要参数进行标定。最后，针对道路横向偏移量M对LSTM-LC模型性能的影响进行对比试验。研究结果表明：相比GRU-LC模型，LSTM-LC模型对换道行为的表征更准确，在模型的精度和迁移性上有着显著的提升；GRU-LC模型的均方差为4.64 m²，迁移性均方差为119.82 m²，而LSTM-LC模型的均方差为3.18 m²，迁移性均方差为79.58 m²，分别优化了31.5%和39.71%；通过引入道路横向偏移量M，可将LSTM-LC模型精度和迁移性提升约10%，且模型稳定性更强。

LSTM-based Lane-changing Behavior Model for Unmanned Vehicle Under Environment of Heterogeneous Human-driven and Autonomous Vehicles

The heterogeneous environment of human-driven and autonomous vehicles (HEHA) is the traffic environment in which human-driven vehicles (HVs) and autonomous vehicles (AVs) run heterogeneously. HEHA requires the driving behavior model of the AV to be similar to that of the HV.In studies of autonomous driving behavior,lane-changing behavior modeling is always a research highlight.According to the theory of deep learning, an LSTM-based lane changing behavior model was constructed for the unmanned vehicle under HEHA.By studying the interaction between HVs and adjacent vehicles during lane changing, the impact factors of the lane-changing behaviors were analyzed. A road lateral offset was introduced into the inputs to improve the accuracy and transferability of the model. An American open traffic dataset, Next Generation SIMulation (NGSIM),was used to build up a lane-changing behavior database by taking the input requirement of LSTM into account.The model used the mean square error (MSE) as the loss function, and the RMS prop optimization method to train. Three vital parameters of the LSTM-the network structure, historical sequence length N, and the training sample size were calibrated. Finally, a comparative experiment was conducted on the impact of road lateral offset on the performance of the LSTM-LC model. The experimental results show that the LSTM-LC model could more accurately characterize the lane-changing behavior, and there was a significant improvement in the accuracy and mobility of the model. The MSE of the GRU-LC model is 4.64 m², and the MSE of migration is 119.82 m². The MSE of the LSTM-LC model is 3.18 m², and the mean square deviation of migration is 79.58 m², which represents decreases of 31.5% and 39.71%, respectively. The introduction of the road lateral offset improves the performance by approximately 10% in terms of the accuracy and transferability of the LSTM-LC model, and the stability of the model training results is stronger.