基于注意力机制及分层网络的危险驾驶行为预测方法

Aggressive Driving Behavior Prediction Method Based on Attention Mechanism and Hierarchical Network

摘要:

基于自然驾驶实验，获取“人?车?环境”多维驾驶行为数据，经过数据清洗与筛选构建危险驾驶行为标准数据库。采用显著性分析对指标进行筛选，并构建八维度的危险驾驶行为预测指标集。以神经网络为第一层，以基于注意力机制的长短期记忆（LSTM）网络为第二层，建立危险驾驶行为预测双层时序模型。结果表明：该模型能有效提升预测准确率（10%）；分层结构和注意力机制对预测准确率有较好的提升作用，分别为5%和3%。

Abstract:

Multi-dimensional driving behavior data， including driver characteristics， road environment， and vehicle operation variables， were collected based on natural driving experiments. Then， the standard database of aggressive driving behavior was created after data cleansing and filtering. The significance analysis was used to obtain useful indexes， with which an eight-dimensional index set for aggressive driving behavior prediction was built. Finally， a two-layer time series model for aggressive driving behavior prediction was constructed. The first layer of the model is an artificial neural network. The second layer of the model is a long short-term memory （LSTM） network with an attention mechanism module. It is shown that the proposed model can increase the prediction accuracy by 10%；the two-layer structure and attention mechanism have a good improvement for the prediction accuracy （5% and 3%， respectively）.