Scholarlink

<p>Road safety conditions are highly dynamic. Even split-second distracted driving behaviors (DDB) can cause disastrous consequences. Therefore, developing an algorithm that can accurately identify DDB in real time is critical for enhancing road safety. To address the issues of excessive model complexity, low detection accuracy, and poor real-time performance in traditional methods, this paper proposes a lightweight and efficient DDB recognition algorithm LEDDR-YOLO based on YOLO11. First, in order to better capture both detailed features of DDB and overall driving posture characteristics, an innovative multi-scale information fusion Neck structure, LMIFFPN, is designed, along with a new feature extraction module, CSP-PGHCB. Second, aiming to reduce model complexity, this paper introduces a more efficient LUP upsampling module and incorporates shared convolution in the detection head, proposing a novel detection head, LSCDetector. Finally, considering the structural characteristics of the LEDDR-YOLO model, a dedicated structured pruning algorithm, LEDDR-Pruning, is developed. This method effectively reduces model size with minimal accuracy loss, and then knowledge distillation strategies are employed to compensate for the accuracy degradation caused by pruning. Experimental results demonstrate that LEDDR-YOLO achieves an accuracy of 99.2% on the StateFarm dataset while maintaining a low computational cost of 1.4 GFLOPs, 404,793 parameters, and 1050 frames per second. Generalization experiments conducted on the 100-Drivers dataset demonstrate that LEDDR-YOLO outperforms YOLO11 in generalization capability. Therefore, the proposed algorithm maintains high accuracy while significantly reducing computational complexity.</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

LEDDR-YOLO: a lightweight and efficient distracted driving recognition algorithm with a particular pruning method

  • 期刊论文

  • 发表日期：

    2025年8月11日

查看全文
  • Qian Shen,
  • Lei Zhang,
  • Yan Zhang,
  • Yuxiang Zhang,
  • Shihao Liu

摘要

Road safety conditions are highly dynamic. Even split-second distracted driving behaviors (DDB) can cause disastrous consequences. Therefore, developing an algorithm that can accurately identify DDB in real time is critical for enhancing road safety. To address the issues of excessive model complexity, low detection accuracy, and poor real-time performance in traditional methods, this paper proposes a lightweight and efficient DDB recognition algorithm LEDDR-YOLO based on YOLO11. First, in order to better capture both detailed features of DDB and overall driving posture characteristics, an innovative multi-scale information fusion Neck structure, LMIFFPN, is designed, along with a new feature extraction module, CSP-PGHCB. Second, aiming to reduce model complexity, this paper introduces a more efficient LUP upsampling module and incorporates shared convolution in the detection head, proposing a novel detection head, LSCDetector. Finally, considering the structural characteristics of the LEDDR-YOLO model, a dedicated structured pruning algorithm, LEDDR-Pruning, is developed. This method effectively reduces model size with minimal accuracy loss, and then knowledge distillation strategies are employed to compensate for the accuracy degradation caused by pruning. Experimental results demonstrate that LEDDR-YOLO achieves an accuracy of 99.2% on the StateFarm dataset while maintaining a low computational cost of 1.4 GFLOPs, 404,793 parameters, and 1050 frames per second. Generalization experiments conducted on the 100-Drivers dataset demonstrate that LEDDR-YOLO outperforms YOLO11 in generalization capability. Therefore, the proposed algorithm maintains high accuracy while significantly reducing computational complexity.