Scholarlink

<p>The World Health Organisation report indicates that despite a slight annual decline in traffic-related fatalities, driver distraction remains the primary cause of traffic accidents, accounting for over 50% of all incidents. Real-time distracted driving detection requires processing continuous video streams with ultra-low latency, a task that fundamentally relies on robust High-Performance Computing (HPC) capabilities. Furthermore, to circumvent the inherent transmission delays of cloud-based supercomputing, the deployment of Edge HPC systems—equipped with highly efficient parallel processing architectures—is imperative. Existing methods suffer from excessive parameters, making them challenging to deploy in vehicle computing. To address this, this paper proposes an innovative architecture—the Binarised Attention-Enhanced Network (BAENet)—with only 0.73&#xa0;M parameters. This network achieves efficient feature learning and enhancement through multi-stage downsampling and feature extraction, combined with attention mechanisms and binarisation techniques. Evaluated on two public datasets—State Farm Distracted Driving Detection (SFD3) and AUC Distracted Driving Detection (AUCD2)—BAENet achieves accuracies of 99.88% and 95.62%, respectively. The model delivers an ultra-fast inference speed of 700.4 FPS on NVIDIA 3090Ti GPU and 121.5 FPS on Jetson AGX Orin, ensuring real-time operation in resource-constrained environments.</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

BAENet: driver distraction detection method based on binarised attention-enhanced network

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Xing Sheng,
  • Jianrong Cao,
  • Junzhe Zhang,
  • Zhen Wang,
  • Zongtao Duan

摘要

The World Health Organisation report indicates that despite a slight annual decline in traffic-related fatalities, driver distraction remains the primary cause of traffic accidents, accounting for over 50% of all incidents. Real-time distracted driving detection requires processing continuous video streams with ultra-low latency, a task that fundamentally relies on robust High-Performance Computing (HPC) capabilities. Furthermore, to circumvent the inherent transmission delays of cloud-based supercomputing, the deployment of Edge HPC systems—equipped with highly efficient parallel processing architectures—is imperative. Existing methods suffer from excessive parameters, making them challenging to deploy in vehicle computing. To address this, this paper proposes an innovative architecture—the Binarised Attention-Enhanced Network (BAENet)—with only 0.73 M parameters. This network achieves efficient feature learning and enhancement through multi-stage downsampling and feature extraction, combined with attention mechanisms and binarisation techniques. Evaluated on two public datasets—State Farm Distracted Driving Detection (SFD3) and AUC Distracted Driving Detection (AUCD2)—BAENet achieves accuracies of 99.88% and 95.62%, respectively. The model delivers an ultra-fast inference speed of 700.4 FPS on NVIDIA 3090Ti GPU and 121.5 FPS on Jetson AGX Orin, ensuring real-time operation in resource-constrained environments.