驾驶人意图识别综述

Review on driver intention recognition

为全面了解驾驶人意图识别研究进展,梳理了近30年关于驾驶人意图识别的研究,将驾驶人意图分类为策略意图、战术意图和操作意图;根据研究热点主要对换道、转向、制动和超车意图进行了归纳;从系统构建的角度对驾驶人意图识别系统的结构、输入、算法和评估进行了综述。根据系统输入的不同,从交通环境、车辆运动以及驾驶人行为对4种驾驶意图进行了综述;并根据构建模型采用算法的不同,从生成模型、判别模型、深度学习、认知模型、基于规则判定以及半监督学习模型6个方面对驾驶意图识别模型的研究进行了综述。结果表明:车辆动态信息在一般情况下不能作为预测驾驶人意图的输入信息,但可作为机动车辆已经开始后检测驾驶人早期意图的有效指标;交通环境和驾驶人行为信号对换道、制动和超车意图的预测非常有效,但是作为转向意图预测的输入并不可靠,车辆行驶轨迹更能反映驾驶人的转向意图;构建不同驾驶人意图识别模型应挑选合适的参数。现有采用机器学习包括深度学习方法构建的驾驶人意图识别模型,存在模型解释性差、对数据样本较为敏感、可扩展性差等局限性;规则判定模型无法适应多变的道路环境和驾驶风格。驾驶人意图识别模型应为自动驾驶技术的发展提供以人为中心的技术支持,能够实现监测驾驶人状态和对交通环境的态势感知,捕捉驾驶人的感知、认知特性,采用半监督学习方法提升模型鲁棒性、减少模型开发时间。在网联交通环境未形成之前,混行网联场景下的驾驶人意图识别模型尚待深入研究。

In order to fully understand the research progress of driver’s intention recognition,the research on driver’s intention recognition in the past 30 years was sorted out.Driver’s intention was classified into strategic intention,tactical intention and operation intention.According to the research hotspots,the intention of changing lanes,turning,braking and overtaking were mainly summarized.The structure,input,algorithm and evaluation of the driver’s intention recognition system from the perspective of system construction was summarized.According to the different input of the system,the four driving intentions were summarized from the traffic environment,vehicle movement and driver behavior.According to the different algorithms used to construct the model,the research on the driving intention recognition model was reviewed from six aspects,such as generative model,discriminant model,deep learning,cognitive model,rule-based decision and semi-supervised learning model.The results show that the vehicle dynamic information can not be used as input information to predict the driver’s intention in general,but it can be used as an effective indicator to detect the driver’s early intention after the vehicle maneuver has started.Traffic environment and driver behavior signals are very effective in predicting lane changing,braking and overtaking intentions,but the input as steering intention prediction is not reliable,and the vehicle trajectory can better reflect the driver’s steering intention.Appropriate parameters should be selected to construct different driver’s intention recognition models.Existing driver intention recognition models constructed using machine learning,including deep learning methods,have limitations such as poor model interpretation,sensitivity to data samples,and poor scalability.The rule judgment model can not adapt to the changing road environment and driving style.The driver’s intention recognition model should provide human-centered technical support for the development of autonomous driving technology.It can monitor the driver’s state and situational awareness of the traffic environment,capture the driver’s perception and cognitive characteristics,and adopt semi-supervised learning methods to improve model robustness and reduce model development time.Before the connected traffic environment is formed,the driver’s intention recognition model in the mixed connected scenario needs to be studied in depth.4 tabs,13 figs,145 refs.