A Machine Learning-Based Approach to Analyze Information Used for Steering Control
A Machine Learning-Based Approach to Analyze Information Used for Steering Control
复制标题
DOI：
10.1109/access.2021.3093337
复制
发表时间：
2021-01-01
期刊：
影响因子：
3.9
通讯作者：
Wada，Takahiro
中科院分区：
文献类型：
--
作者：
Okafuji，Yuki;Sugiura，Toshihito;Wada，Takahiro
理解驾驶行为与视觉信息之间的关系对于全面理解驾驶行为具有重要意义。然而，转向/油门控制的认知行为分析仅在特殊的模拟器环境下进行。因此，在这项研究中，我们的目标是开发一个具有人类物理特征的卷积神经网络（CNN）来分析驾驶员的认知行为，并验证机器学习方法可以成为理解驾驶员行为的分析方法。我们在模拟器实验中获得了驾驶数据来训练所提出的CNN模型。使用由训练的CNN模型生成的特征图的结果和驾驶员的注视行为来分析视野影响驾驶员的转向行为的区域。结果表明，驾驶员使用距注视点20度内的信息来执行转向控制。这表明，从我们提出的方法获得的结果可以重现与以前的发现相同的结果。我们还验证了结果不是唯一获得的，取决于所提出的模型和环境，但也受到驾驶行为，如凝视点和转向控制。我们分析了由数学控制模型生成的数据集，该模型称为驾驶员模型，它执行与驾驶员不同的行为。驾驶员模型产生的分析结果与人类数据的结果不同。因此，基于机器学习的分析生成的结果受到驾驶行为的影响。因此，这些结果意味着机器学习方法有可能成为理解驾驶员行为的分析方法。
Understanding the relationship between driving behavior and visual information is important for holistic understanding of driving behavior. However, the analysis of the cognitive behavior for steering/throttle control has been only conducted under a special simulator environment. Therefore, in this study, we aimed to develop a convolutional neural network (CNN) with human physical characteristics to analyze the driver's cognitive behavior and to validate that the machine learning methods can be an analytical method for understanding driver behavior. We obtained the driving data in a simulator experiment to train the proposed CNN model. The region where the visual field influences drivers' steering behavior was analyzed using the results of the feature maps generated by the trained CNN model and the driver's gaze behavior. The results indicate that the driver performs steering control using the information within 20 degrees from the gaze point. This shows that the results obtained from our proposed method can reproduce the same results as previous findings. We also validated that the results are not uniquely obtained depending on the proposed model and environment but are also influenced by the driving behavior such as the gaze point and the steering control. We analyzed the dataset generated by the mathematical control model, called the driver model, which performs different behaviors from the driver. The analysis results generated by the driver model were different from the results of the human data. Therefore, the results generated by the machine learning-based analysis are influenced by the driving behavior. Consequently, these results imply that machine learning methods have the potential to become analytical methods for understanding driver behavior.