基于迁移学习和深度学习的驾驶员分心行为识别研究

Driver distraction behavior recognition based on transfer learning and deep learning

为了解决传统驾驶员分心行为识别模型准确率过度依赖大样本数据集、耗时较长等问题,提出了一种结合迁移学习策略和卷积神经网络模型的方法来对驾驶员分心行为进行识别。首先在模型中引入ImageNet数据集上训练好的网络权重,冻结网络的卷积层;然后去掉原网络中的全连接层,重新添加输出维度为10的FC层;最后在验证集上对比基于迁移学习策略模型与原网络模型的识别精度。结果表明,基于迁移学习策略的分心行为识别模型比原网络模型的平均准确率提升了约4%,显著提高了分心行为的识别率。本研究结果可为驾驶员分心行为识别提供理论与技术支持。

In order to solve the problem of traditional recognition model of drivers’distracted behavior,including relying on large amounts of data sets and time-consuming,this paper proposes a method combining transfer learning strategy and convolutional neural network model to conduct recognition research on drivers’distracted behavior.Firstly,the weight of network trained on the ImageNet data set was introduced into the model,and the convolutional layer was frozen.Then,the full connection layer in the original network was removed,and the FC layer with output dimension of 10 was added.Finally,the recognition accuracy based on the migration learning strategy was compared with that of the original network model on the verification set.The results show that the average prediction accuracy of the recognition model based on transfer learning strategy is improved by about 4%compared with the original network model,and the recognition rate of distracted behavior is significantly improved,which can provide theoretical and technical support for drivers’distracted behavior recognition.