基于局部特征的卷积神经网络车灯识别

Recognition of Vehicle Lights Based on Convolutional Neural Network with Local Features

为了解决车辆管控工作中出现的肇事车辆逃避交通监管的问题,对数据集处理方式和局部特征的车型分类算法进行研究。首先,以AlexNet网络为基础分析了各个网络结构对于输入图片的敏感程度,从网络层数和卷积核尺寸上进行网络优化得出IM-AlexNet网络。然后,使用数据增强方式处理后的自建数据集,训练IM-AlexNet分类模型网络。最后,在HOG-SVM、GoogleNet和VGG16三种模型上进行对比实验并分析。实验结果表明:IM-AlexNet网络在验证集上准确率达到96%左右,损失值低于0.2,训练速度达到3 s/step。在混淆矩阵中IM-AlexNet网络模型总体准确率达到69%,完成了局部特征对车型分类的实验,分类准确率大大提高。

In order to solve the problem that the vehicle causing the accident evades the traffic supervision in the vehicle control work,the processing method of data set and the vehicle classification algorithm of local features are studied.Firstly,the sensitivity of each network structure to the input image is analyzed based on the AlexNet network.The IM-AlexNet network is obtained from the network layer number and the size of the convolution kernel.Then,the self-built data set processed by data enhancement method is used to train the IM-AlexNet classification model network.Finally,comparative experiments and analysis are carried out on HOG-SVM,GoogleNet and VGG16 models.The experimental results show that the accuracy of IM-AlexNet network on the verification set is about 96%;the loss value is less than 0.2;and the training speed is up to 3 s/step.In the confusion matrix,the overall accuracy of IM-AlexNet network model reached 69%;the classification experiment of local features is completed;the classification accuracy is greatly improved.