基于卷积神经网络(CNN)的高速路交通图片拥堵识别

Congestion Identification of Traffic Photos Based on Convolutional Neural Networks(CNN)

近年来,高速路的拥堵问题变得越来越严重,传统的交通拥堵识别采用视频进行研究,其具有代价昂贵,识别速度慢的缺点。文章提出了一种基于图片进行交通拥堵识别的方法。因为卷积神经网络(CNN)在图像识别方面有着识别速度快,适用范围广,识别准确率高的优点,所以文章使用了带有relu激活函数代替传统的sigmod函数和tanh函数,并引入了dropout层的卷积神经网络模型Google Net,并对网络结构和参数进行了调整优化,得到了一个交通拥堵图片识别的改进的Google Net改进模型,该模型的样本内测试准确率达到了98.6%。在对2000张现实高速路上的图片进行识别测试后,测得其准确率为96.5%。采用文理特征的传统方法的高速路交通拥堵图像识别准确率为90%。

In recent years,the problem of highway congestion has become more and more serious. Traditional traffic congestion identification is studied by video,which has the disadvantages of high cost and low recognition speed. This paper presents a method of traffic congestion identification based on pictures. Because the convolution neural network（CNN） has the advantages of high recognition speed,wide range of application and high recognition accuracy in image recognition,we use the Re LU activation function instead of the traditional Sigmod function and Tanh function, and introduce the Dropout layer convolution neural network model Google Net. The network structure and parameters are adjusted and optimized,and an improved Google Net model for traffic congestion image recognition is obtained. The test accuracy of the model is 98.6%. After the recognition test of 2,000 images on the real highway, the accuracy rate is 96.5%. The recognition accuracy of highway traffic congestion image using traditional methods of literary and scientific features is 90%.