基于多头注意力的CNN-LSTM的换道意图预测

自动驾驶车辆与传统车辆混行的交通环境中,车辆的换道意图预测能够为自动驾驶车辆安全行驶提供有效保证。为了更准确地预测车辆的换道意图,将多头注意力与卷积神经网络(Convolution neural network,CNN)和长短时记忆(Long-short term memory,LSTM)网络结合,提出一种新型车辆换道意图预测算法。首先对NGSIM(Next generation Simulaion)数据集进行处理,提取车辆横向位置信息和周围环境信息。然后输入基于多头注意力(Multi-headattention)的CNN-LSTM模型,提高对输入序列特征的提取能力和预测精度。最后在NGSIM数据集验证该模型的有效性。试验结果表明,该模型能够从大量数据中提取到重要特征,同时通过特征对比试验发现,横向位置信息作为预测的主要特征,而周围环境信息作为预测的辅助特征。最后通过模型的对比试验得出,该模型的换道意图预测准确率在换道前1s、2s、3s相比于LSTM、CNN、CNN-LSTM模型具有更好的预测精度,可以为自动驾驶汽车设计先进的意图预测算法提供帮助和参考。

Lane Change Intention Prediction of CNN-LSTM Based on Multi-head Attention

In the mixed traffic environment of autonomous vehicles and traditional vehicles, the prediction of vehicle lane change intention can provide an effective guarantee for the safe driving of autonomous vehicles. In order to more accurately predict the vehicle lane change intention, a novel vehicle lane change intention prediction algorithm is proposed by combining the multi head attention with convolution neural network(CNN) and long-short term memory(LSTM) network. Firstly, the NGSIM(Next generation simulaion) data set is processed to extract the vehicle lateral position information and surrounding environment information. Then, it is input into the CNN-LSTM model based on multi-head attention to improve the feature extraction ability and prediction accuracy of the input sequence. Finally, the effectiveness of the proposed model is verified on the NGSIM dataset. The experimental results show that the model can extract important features from a large number of data. At the same time, through the feature comparison experiment, it is found that the lateral position information is the main feature of prediction, and the surrounding environment information is the auxiliary feature of prediction. Finally, through the comparison experiment of the model, it is concluded that the prediction accuracy of the model is better than that of LSTM, CNN and CNN-LSTM models before 1s, 2s and 3s. It can provide help and reference for the design of advanced prediction algorithms for autopilot cars.