Abstract: In order to improve the accuracy of vehicle driving style recognition, the article builds a driving style recognition method based on long and short-term memory neural network combined with hybrid schematic neural network. The article dissects the relationship between drivers' driving styles and time series data by mining the time series features in driving behavior data, using the public dataset NGSIM, extracting descriptive features after screening and smoothing of the dataset, converting the high-dimensional features to low-dimensional ones by principal component analysis, and specifying the number of clusters by K-means method to be identified as three driving styles. Then the classification recognition of driving styles is carried out by multi-task learning, multi-task fusion, long short-term memory (MTL-MTF-LSTM) neural network, which combines reinforcement learning and imitation learning, after multi-task assignment of recognition pool. The results show that the model achieves 95%, 98% and 97% recognition accuracy for conservative, general and aggressive driving styles, respectively, with an overall excellent performance.

Key words: driving style recognition; reinforcement learning; hybrid teaching neural network; multi-task learning; cluster analysis

摘要： 为了提高车辆驾驶风格的辨识准确度，文章基于长短期记忆神经网络结合混合示型神 经网络搭建了一种驾驶风格识别方法。通过挖掘驾驶行为数据中时序特点，剖析驾驶人驾驶 风格与时序数据的关系，使用公开数据集 NGSIM，经过数据集的筛选与平滑处理后提取描述 性特征，采用主成分分析法转换高维特征为低维，通过 K-means 方法指定聚类数量，确定为 三种驾驶风格。通过多任务学习多任务融合长短期记忆（MTL-MTF-LSTM）神经网络进行了 驾驶风格的分类识别，该模型结合了强化学习和模仿学习，经过多任务分配识别池。结果表 明模型对保守型、一般型和激进型驾驶风格的识别精度分别达到了 95%、98%和 97%，整体 表现优异。

关键词: 驾驶风格识别；强化学习；混合示教型神经网络；多任务学习；聚类分析