轨迹数据驱动的车辆换道意图识别研究

为实现准确识别车辆换道意图，提高车辆行驶安全性，综合考虑车辆换道过程的时空特性及不同特征对车辆的影响程度，提出一种基于卷积神经网络(CNN)与门控循环神经网络(GRU)组合并融合注意力机制的换道意图识别模型。首先，筛选和平滑处理车辆轨迹数据，将车辆轨迹数据分为向左换道、向右换道及直线行驶3类，构建换道意图样本集。其次，构建融合注意力机制的 CNN_GRU模型，识别换道意图样本集，考虑到行驶过程中车辆之间的交互性，将被预测车辆和周围车辆的位置和速度信息作为模型的输入，经过CNN层特征提取的特征作为GRU层的输入，经过注意力机制层对不同的特征增加不同的权重系数，利用 Softmax 层识别换道意图。最后，选用 NGSIM 中 US-101 数据集的轨迹数据验证融合注意力机制的 CNN_GRU模型性能， 同时，与LSTM、GRU、CNN_GRU及CNN_LSTM_Att等模型进行对比分析。验证结果表明，所提模型车辆换道意图识别整体准确率达到97.37%，迭代时间为6.66 s，相比于其他模型准确率最多提高9.89%，最少提高2.1%。分析不同预判时间下的意图识别，模型可在车辆换道前2 s 内均能识别换道意图，准确率在89%以上，表现出良好的识别性能。

Vehicle Lane Change Intention Recognition Driven by Trajectory Data

In order to accurately identify the vehicle's lane-changing intention and improve the driving safety of the vehicle, I comprehensively considered the spatiotemporal characteristics of the vehicle's lane-changing process and the influence of different characteristics on the vehicle, and proposed a lane-changing intention recognition model with attention mechanism, which is based on the combination of Convolutional Neural Network (CNN) and Gated Recurrent Unit Neural Network (GRU). Firstly, I filtered and smoothed the vehicle trajectory data, and divided the vehicle trajectory data into three categories: left lane change, right lane change, and straight driving, so as to construct a sample set of lane change intention. Secondly, I built a CNN_GRU model that integrates attention mechanism to identify the sample set of lane change intention. Considering the interaction between vehicles during driving, I utilized the position, the speed information of the predicted vehicle and surrounding vehicles as the input of the model. After the CNN layer feature extraction, I then chose the extracted features as the input of GRU layer. And I also added different weight coefficients to different features through the attention mechanism layer, and leveraged the Softmax layer to identify the lane change intention. Finally, I verified the performance of CNN_GRU model with fused attention mechanism by using the trajectory data of US-101 dataset in NGSIM, and at the same time, compared and analyzed it with LSTM, GRU, CNN_GRU and CNN_LSTM_Att models. The results showed that the proposed model achieves an overallaccuracy of 97.37% for vehicle lane change intention recognition with an iteration time of 6.66 s, which is at most 9.89% and at least 2.1% improvement in accuracy compared to other models. By analyzing the accuracy of intention recognition at different pre-determination times, we know that the intention to change lanes can be accurately recognized within 2 s before the vehicle changes lanes, and the accuracy rate is above 89%, so the model has good recognition performance.