Multimodel Approach to Personalized Autonomous Adaptive Cruise Control
Multimodel Approach to Personalized Autonomous Adaptive Cruise Control
复制标题
DOI：
10.1109/tiv.2019.2904419
复制
发表时间：
2019-06-01
影响因子：
8.2
通讯作者：
Jia, Yunyi
中科院分区：
文献类型：
--
作者：
Bolduc, Andrew Phillip;Guo, Longxiang;Jia, Yunyi
自动驾驶汽车正受到越来越多的关注，但调查显示，很大一部分人对采用这项新技术持谨慎态度。犹豫不决的一个可能解释是，由于制造商设置的控制模型及其参数，乘员不会对驾驶风格感到舒服。舒适度本质上是主观的，因此因人而异。为了解决这个问题，自动驾驶汽车必须能够适应用户的驾驶风格偏好。如果我们假设司机对自己的驾驶风格更满意，我们可以选择让车辆学习，并将司机的风格纳入控制模型。然而，目前还没有被广泛接受的“最佳”模式。一种车型可能被证明比其他车型更能代表特定的司机，尽管司机可能会选择不安全的驾驶条件，这种情况的复制不应优先于乘员的安全。在本文中，我们提出了一种多模型方法来寻找描述个人在高速公路上的纵向驾驶风格的最佳驾驶员模型。首先提出了一种个体驾驶风格指标的提取方法。然后，详细描述了一种基于多模型的评价方法。本文对Chandler、Herman和Montroll、通用汽车的非线性模型、Tampere模型、Addison&Low模型、最优速度模型和神经网络模型进行了训练和比较。在复制驾驶风格方面具有最好性能的模型进一步与模型预测控制器相耦合，以包括安全驾驶的安全约束。最后，使用从五个不同驾驶员那里收集的驾驶数据对所提出的多模型方法进行了测试。测试结果表明，我们的基于多模型的方法在模仿个人的纵向驾驶风格方面比单一模型方法显示出优势。
Autonomous vehicles are gaining increased attention but surveys have shown that a large percentage of people are wary of adopting the new technology. One possible explanation for the hesitancy is that the occupant would not be comfortable with the driving style as a result of the control models and their parameters as set by the manufacture. Comfort level is subjective in nature and therefore varies between individuals. To combat this issue, autonomous vehicles must be able to adapt to the driving style preference of the user. If we assume drivers are more comfortable with their own driving style, we can choose to have the vehicle learn and incorporate the driver's style into the control models. However, there is still no widely accepted "best" model. One modelmay prove to better represent a particular driver than other models though a driver may choose unsafe driving conditions, the replication of which should not take precedence over the safety of the occupant. In this paper, we propose a multimodel approach to find the best driver model for describing an individual's longitudinal driving style on highway. A method for extracting the indicators of an individual's driving style is proposed first. Then, a multimodel-based evaluation method is described in detail. Chandler, Herman,& Montroll, General Motors Nonlinear, Tampere, Addison & Low, Optimal Velocity Model, and Neural Network models are trained and compared in this paper. The model with the best performance in replicating driving style is further coupled with a model predictive controller to include safety constraints for safer driving. Finally, the proposed multimodel approach is tested with driving data collected from five different drivers. The test results show that our multimodel-based approach is showing advantage over a single model approach in imitating an individual's longitudinal driving style.