摘要：

由驾驶员的不安全行为导致的交通事故占多数, 针对驾驶认知素质特性的研究, 搭建虚拟驾驶场景评估驾驶者的驾驶素质, 可以最大限度地贴近现实环境和操作, 唤醒驾驶者的潜在驾驶能力和应对能力。眼球运动可以极大程度地反映出驾驶者的认知状态, 但目前多数眼动状态识别研究主要关注在自然状态中基本视觉运动方向或者眼睑的闭合, 识别类别的能力和效果对于驾驶场景的认知状态评估有限。收集了10类静态眼动方向的双眼数据, 并提出融合注意力机制的多尺度眼状态图像识别模型。首先, 使用部分卷积设计双分支特征融合模块, 在加强模型特征提取能力的同时减少计算冗余; 然后, 在双分支特征融合的残差模块中嵌入改进的坐标注意力(CA)机制, 提升模型对不同尺度特征的信息表征能力; 最后, 对模型的通道结构和数量进行调整, 平衡模型的参数量与识别准确率。实验结果表明, 所提方法在构建的10类眼动状态数据集上识别准确率达到95.1%, 相比改进前的网络提高3.4个百分点; 在Eye Chimera数据集和MRL眼睛数据集上的识别准确率分别为95.1%和98.95%, 可以满足在虚拟驾驶测试环境下眼动状态识别的要求, 并为进一步结合多参数分析驾驶素质缺失任务奠定基础。

关键词: 驾驶认知, 眼状态, 图像分类, 特征融合, 注意力模块

Abstract:

Unsafe driver behaviors account for the majority of traffic accidents. For the purposes of investigating driver cognitive quality characteristics, virtual driving scenes are constructed to mimic the real environment and operation. This helps assess the driver's driving and coping abilities and is of positive significance in reducing road killers. Eye movements can significantly reflect a driver's cognitive state. However, most current eye movement state recognition studies primarily focus on basic visual motion direction or eyelid closure in the natural state, and the ability and effect of recognizing the categories are limited with regard to the cognitive state assessment of driving scenarios. This study collects binocular data from ten categories of static eye movement directions and proposes a multi-scale eye state image recognition model incorporating the attention mechanism. First, a two-branch feature fusion module is designed using partial convolution to enhance the feature extraction capability of the model while reducing computational redundancy. Second, an improved Coordinate Attention(CA) mechanism is embedded in the two-branch feature fusion residual module, to enhance the model's ability to characterize feature information at different scales. Finally, the structure and number of channels are adjusted to balance recognition accuracy and the number of parameters in the model. Experimental results demonstrate that the proposed method achieves a recognition accuracy of 95.1% on the proposed 10-class eye movement state dataset, which is 3.4 percentage points higher than that of the pre-improvement network. The recognition accuracy on the Eye Chimera and MRL eye datasets is 95.1% and 98.95%, respectively, which satisfies the eye movement state recognition requirements for virtual driving test environments, and lays the foundation to further combine multi-parameter analysis for driving quality deficiency tasks.

Key words: driving cognition, eye status, image classification, feature fusion, attention module