面向前车的驾驶行为感知与意图识别算法研究

感知周围车辆的驾驶行为并识别其意图将成为新一代高级驾驶辅助系统的重要组成部分。针对现有方法只考虑单一驾驶行为且可扩展性和可伸缩性差,提出一种基于稀疏表示理论的驾驶行为感知字典模型(Driving Behavior Perception Dictionary Model, DBPDM)。将车辆行驶状态视为时间序列,设计基于自回归积分移动平均(Autoregressive Integrated Moving Average,ARIMA)结合在线梯度下降(Online Gradient Descent, OGD)优化器的在线预测模型,提出基于驾驶行为预测的意图识别构架(Intention Recognition Framework, IRF)。首先,采用图Lasso方法估计典型驾驶行为的稀疏逆协方差矩阵构建驾驶行为字典库,并采用Logdet散度方法计算各逆协方差矩阵的差异获得行为感知字典模型。然后,基于在线预测模型对目标车辆的行驶轨迹和运动状态进行预测,结合主车车辆的行驶状态作为稀疏表示的观测信号,以获取预测时域内的目标车辆意图。最后,采用NGSIM (Next Generation SIMulation)真实驾驶数据对模型进行开发和测试。研究结果表明:所提出的行为感知模型能对6种典型驾驶行为构建行为字典,在分类准确率上与现有方法相比有明显提升,对换道和转向行为样本的平均识别准确率分别达到99.1%和92.9%;该模型能够在相对早期阶段准确地识别出车辆行为;在线预测算法能较好预测出目标车辆的行驶轨迹和运动状态,从而间接地反映出其在预测时域内的驾驶意图;IRF可在换道和转向行为开始前的1.5 s较为准确地识别出目标车辆的意图,平均识别准确率超过80%。

Research on Algorithms of Driving Behavior Perception and Intention Recognition Oriented to the Vehicle in Front

Perceiving the driving behavior of surrounding vehicles and recognizing their intentions will become an important part of next-generation advanced driver assistance systems. To address the problem that existing modeling methods only consider a single driving behavior and have poor extensibility and scalability, a driving behavior perception dictionary model (DBPDM) based on sparse representation theory was proposed. In addition, by regarding the vehicle driving state as a time series, an online prediction model based on an autoregressive integral moving average (ARIMA) model combined with an online gradient descent (OGD) optimizer was designed, and an intention recognition framework (IRF) based on driving behavior prediction was proposed. First, the graphical Lasso method was used to estimate the sparse inverse covariance matrix of typical driving behavior to construct a driving behavior dictionary library, and the Logdet divergence method was used to calculate the difference of each inverse covariance matrix to obtain the behavior perception dictionary model. Then, based on the online prediction model, the driving trajectory and motion state of the target vehicle were predicted, and the driving state of the host vehicle was used as a sparsely represented observation signal to obtain the intention of the target vehicle in the prediction horizon. Finally, the model was developed and tested using Next Generation Simulation (NGSIM) real driving data. The experimental results show that the proposed behavior perception model can construct a behavior dictionary for six typical driving behaviors. The classification accuracy of the proposed method is substantially higher than that of existing methods, and the average recognition accuracy for lane changing and turning behavior samples reached 99.1% and 92.9%, respectively, indicating that the model can accurately recognize vehicle behavior at a relatively early stage. Thus, the model can predict the driving trajectory and motion state of a target vehicle, thereby indirectly reflecting its driving intention in the prediction horizon. The IRF can accurately identify the intention of the target vehicle 1.5 s before the start of lane changing and turning behavior, with an average recognition accuracy of over 80%.