Scholarlink

<p>A novel deep learning framework for recognition of distracted driving behavior is proposed in this paper. The proposed framework consists of hybrid convolutional neural network and bidirectional long short term memory network to extract multi-scale spatiotemporal features of high-dimensional distracted behavior data. The attention mechanism is utilized to assign higher weight coefficients to critical features so as to reduce the impact of redundant information. A fully connected neural network layer is applied to establish a nonlinear mapping relationship between the extracted features and driving behavior categories. Bayesian optimization algorithm is adopted to automatically optimize hyperparameters of the network so as to improve training efficiency and performance of the proposed model. The experimental results demonstrate that the proposed method get an accuracy of about 99.75% on the publicly dataset StateFarm, and outperform the other models in term of recognition accuracy.</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

A novel method for distracted driving behaviors recognition with hybrid CNN-BiLSTM-AM model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月23日

查看全文
  • Dengfeng Zhao,
  • Haojie Li,
  • Zhijun Fu,
  • Bao Ma,
  • Fang Zhou,
  • Chaohui Liu,
  • Wenbin He

摘要

A novel deep learning framework for recognition of distracted driving behavior is proposed in this paper. The proposed framework consists of hybrid convolutional neural network and bidirectional long short term memory network to extract multi-scale spatiotemporal features of high-dimensional distracted behavior data. The attention mechanism is utilized to assign higher weight coefficients to critical features so as to reduce the impact of redundant information. A fully connected neural network layer is applied to establish a nonlinear mapping relationship between the extracted features and driving behavior categories. Bayesian optimization algorithm is adopted to automatically optimize hyperparameters of the network so as to improve training efficiency and performance of the proposed model. The experimental results demonstrate that the proposed method get an accuracy of about 99.75% on the publicly dataset StateFarm, and outperform the other models in term of recognition accuracy.