基于长短期记忆网络的换道意图识别方法

准确识别周围车辆的换道意图将有助于自动驾驶系统决策，从而提升安全性和舒适性。提出一种基于长短期记忆（Long Short-term Memory，LSTM）网络的换道意图识别方法，能够较为准确地识别周围车辆的换道意图。该方法先通过构造收益函数来描述目标车辆（被预测的车辆）与其邻域车辆之间的交互关系，得到目标车辆左换道、右换道和车道保持的收益值，并将该收益值作为交互特征输入到意图识别网络；在意图识别网络中，引入注意力机制，通过网络自学习得到的权重对LSTM层各个时刻的输出加权求和，能够对编码信息进行有效利用，提高换道意图的识别性能；由于车辆的换道意图存在较强的前后依赖性，引入条件随机场（Conditional Random Field，CRF），采用意图转移特征函数对各个时刻换道意图进行联合建模，并构建负对数似然损失函数作为整个网络的损失。为了验证所提方法的有效性，基于NGSIM数据集训练并评估模型。结果表明：所提方法对换道意图识别的准确率、宏观F1分数、测试集损失分别为0.916 4、0.874 6和0.168 3，均优于支持向量机（SVM）、隐马尔可夫模型（HMM）和LSTM模型。同时，所提模型对左换道和右换道的平均换道提前识别时间分别为3.08、2.33 s，综合换道提前识别时间为2.81 s，优于基线模型，能够为主车的决策提供充足的冗余时间。通过消融分析可知，引入的交互作用模块、注意力机制和条件随机场对准确率的贡献分别为0.012 2、0.004 3和0.011 0，印证了相关模块的有效性。最后由场景验证的案例可以得出，所提方法在准确率、稳定性和换道提前识别时间等指标上优于对比模型。

Lane Change Intention Recognition Method Based on an LSTM Network

Accurate assessment of the lane-changing intentions of surrounding vehicles will benefit the decision-making of autonomous driving systems, toward enhanced driving safety and comfort. We propose herein a novel method to recognize accurately the lane-change intentions of a target vehicle based on long short-term memory (LSTM). This method models the interaction between the target and adjacent vehicles via construction of utility functions, followed by the utility values of the target vehicle being obtained for the intention of executing left-lane-change (LLC), right-lane-change (RLC), and lane-keeping (LK) maneuvers. These values are then input into the lane-changing intentions recognition network as interaction features. In addition, the attention mechanism was introduced to the network, and the weights learned by the network were used to sum the weighted outputs of the LSTM layer at each time step, which was expected to use the encoding information effectively and improve the recognition performance. Owing to the strong dependence between lane-changing intentions at different time steps (the intention of current time step depends on the last time step's intention to a certain extent), we introduced the conditional random field (CRF), which jointly and continuously models such intentions by the transfer characteristic function and constructed a negative logarithmic likelihood loss function for the entire network. To verify the validity of the method in this study, we trained and validated the model based on the NGSIM dataset, and the results show that the accuracy, macro F1 score, and test dataset loss of the proposed model are 0.916 4, 0.874 6, and 0.168 3, respectively, which are better than those of the support vector machine (SVM), hidden Markov model (HMM), and LSTM model. Currently, the proposed model recognizes LLC and RLC intentions in average times of 3.08 s and 2.33 s in advance respectively while the corresponding comprehensive time is 2.81 s, which is better than the baseline models. It also provides sufficient time redundancy allowing for the subject vehicle's decision. Through ablation analysis, the contributions of the interaction module, attention mechanism, and CRF to the accuracy rate are 0.012 2, 0.004 3, and 0.011 0, respectively, demonstrating their efficacy. Finally, in scenario validation, the method proposed in this paper is superior to the baseline models regarding accuracy, stability, and advance lane-changing recognition time.