Scholarlink

<p>With the continuous development of intelligent transportation systems, driving safety has become a core societal concern. Distracted driving, one of the leading factors of traffic accidents, has been directly responsible for over 60% of such incidents. To effectively reduce these accidents, real-time monitoring and early warning of distracted driving have become crucial tasks in enhancing traffic safety. In this paper, we proposes a low-latency, high-efficiency network called USH, aiming to quickly and accurately detect driver distraction behavior on low-computation devices. USH employs a stage-wise hybrid modeling approach, introducing convolutional operations and self-attention at different stages of the model to achieve a balance between model accuracy and efficiency. By injecting local information modeling layers into a Feed Forward Network (FFN), the network complexity is significantly reduced while maintaining high detection accuracy. Additionally, USH introduces a single-head self-attention, effectively avoiding the head redundancy of multi-head self-attention, making the model more lightweight and efficient, meeting the detection needs in resource-constrained environments. Experimental results demonstrate that USH achieves a top-1 accuracy of 98.5% on the StateFarm dateset, with an inference latency of only 4.7 ms. Compared to the original model, the proposed USH achieves a 21.6% improvement in computational efficiency, a 146% reduction in model size. These results indicate that USH not only demonstrates excellent performance and stability, but also exhibits significant potential in the field of distracted driving detection.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

USH: an efficient real-time distracted driving detection model

  • 期刊论文

  • 发表日期：

    2025年8月14日

查看全文
  • He Wang,
  • Yuan Li

摘要

With the continuous development of intelligent transportation systems, driving safety has become a core societal concern. Distracted driving, one of the leading factors of traffic accidents, has been directly responsible for over 60% of such incidents. To effectively reduce these accidents, real-time monitoring and early warning of distracted driving have become crucial tasks in enhancing traffic safety. In this paper, we proposes a low-latency, high-efficiency network called USH, aiming to quickly and accurately detect driver distraction behavior on low-computation devices. USH employs a stage-wise hybrid modeling approach, introducing convolutional operations and self-attention at different stages of the model to achieve a balance between model accuracy and efficiency. By injecting local information modeling layers into a Feed Forward Network (FFN), the network complexity is significantly reduced while maintaining high detection accuracy. Additionally, USH introduces a single-head self-attention, effectively avoiding the head redundancy of multi-head self-attention, making the model more lightweight and efficient, meeting the detection needs in resource-constrained environments. Experimental results demonstrate that USH achieves a top-1 accuracy of 98.5% on the StateFarm dateset, with an inference latency of only 4.7 ms. Compared to the original model, the proposed USH achieves a 21.6% improvement in computational efficiency, a 146% reduction in model size. These results indicate that USH not only demonstrates excellent performance and stability, but also exhibits significant potential in the field of distracted driving detection.