Zhang Lianlong, Chen Xiaodong, Chen Zexin, Zheng Jiawen, Diao Yinliang
College of Electronic Engineering, South China Agricultural University, Guangzhou 510642, China.
Sensors (Basel). 2025 Apr 10;25(8):2399. doi: 10.3390/s25082399.
Driver status plays a critical role in ensuring driving safety. However, the current visual recognition-based methods for detecting driver actions and status are often limited to factors such as ambient light condition, occlusion, and privacy concerns. In contrast, millimeter-wave radar offers various advantages such as high accuracy, ease of integration, insensitivity to light condition, and low cost; therefore, it has been widely used for monitoring vital signals and in action recognition. Despite this, the existing studies on driver action recognition have been hindered by limited accuracy and a narrow range of detectable actions. In this study, we utilized a 77 GHz millimeter-wave frequency-modulated continuous-wave radar to construct a dataset encompassing seven types of driver head-hand cooperative actions. Furthermore, a deep learning network model based on VGG16-LSTM-CBAM using micro-Doppler spectrograms as input was developed for action classification. The experimental results demonstrated that, compared to the existing CNN-LSTM and ALEXNET-LSTM networks, the proposed network achieves a classification accuracy of 99.16%, effectively improving driver action detection.
驾驶员状态在确保驾驶安全方面起着关键作用。然而，当前基于视觉识别的检测驾驶员行为和状态的方法通常受到环境光条件、遮挡和隐私问题等因素的限制。相比之下，毫米波雷达具有高精度、易于集成、对光照条件不敏感和低成本等多种优点；因此，它已被广泛用于监测生命体征和动作识别。尽管如此，现有的驾驶员动作识别研究受到精度有限和可检测动作范围狭窄的阻碍。在本研究中，我们利用77 GHz毫米波调频连续波雷达构建了一个包含七种驾驶员头部-手部协同动作的数据集。此外，还开发了一种基于VGG16-LSTM-CBAM的深度学习网络模型，以微多普勒频谱图作为输入进行动作分类。实验结果表明，与现有的CNN-LSTM和ALEXNET-LSTM网络相比，所提出的网络实现了99.16%的分类准确率，有效提高了驾驶员动作检测能力。
