基于信息融合的多列卷积神经网络异常驾驶研究

Research on abnormal driving in multi⁃column convolutional neural networks based on information fusion

针对传统卷积神经网络(CNN)在异常驾驶检测方面识别率低、处理能力差的问题。对异常驾驶状态进行研究,提出了一种基于信息融合的多列卷积神经网络用来检测异常驾驶状态。首先建立三个卷积局部感受野核大小不同的CNN,然后将三列卷积神经网络卷积特征图进行融合,最后,融合的卷积特征图通过全连接层进行降维,输出不同的驾驶行为。实验结果表明,该方法相对于传统的卷积神经网络取得了更高的准确性、鲁棒性。在state farm distracted driver detection数据集上识别率达到了89.8%。

The traditional convolutional neural network(CNN)has a low recognition rate and poor processing ability in abnormal driving detection methods.The abnormal driving state is studied,and a multi‑column convolutional neural network based on information fusion is proposed to detect the abnormal driving state.First,three CNNs with different sizes of convolutional local Receptive field nuclei are established,and the three‑column convolutional neural network convolution feature map is fused.Finally,the fused convolutional feature map is dimensionally reduced through the full connection layer,and different driving behaviors are output.The analysis of experimental results shows that this method achieves higher accuracy and robustness compared to traditional convolutional neural networks.The recognition rate on the state farm dispersed driver detection dataset reached 89.8%.