北京工业职业技术学院学报

针对Transformer网络应用于车载嵌入式平台时面临的算力与速度瓶颈问题，以实现高效、轻量的分心驾驶行为识别为目的，设计了一种能够精准聚焦图像关键区域、有效提升特征提取效率的轻量化视觉网络CR-Transformer。实验结果显示：与近几年先进的视觉网络相比，在分心驾驶行为识别任务中，CR-Transformer的识别准确率更高，浮点运算次数和参数规模也更少，可为硬件资源受限的分类任务提供新的思路与技术参考。

To address the computational power and speed bottlenecks of Transformer networks applied to in-vehicle embedded platforms, this paper designs a lightweight visual network named CR-Transformer to achieve efficient and lightweight distracted driving behavior recognition. The proposed network can precisely focus on key image regions and effectively improve feature extraction efficiency. Experimental results show that, compared with state-of-the-art visual networks in recent years, CR-Transformer achieves higher recognition accuracy, fewer floating-point operations and smaller parameter size in the distracted driving behavior recognition task. It provides new ideas and technical references for classification tasks under hardware resource constraints.