Highway Congestion Recognition Based On Deep Convolutional Neural Networks

In recent years,the traffic congestion has become more and more serious,causing inestimable traffic safety threats,environmental pollution,and severe economic losses.Therefore,it is urgent to alleviate the congestion problem.Real-time detection of highway congestion is convenient for transportation departments to monitor,which is an important means to effectively alleviate congestion problems.Traffic jam recognition algorithms can rely on coils,GPS and other methods.Today,many cities are installed a large number of camera networks,the installation cost is low,and you can learn more about traffic flow information.These surveillance videos can be used to detect traffic congestion by calculating the traffic flow density,but this method is easily affected by the quality of video transmission,and has a large amount of calculation and a long detection delay.Nowadays,hardware conditions such as high-performance computers and GPU graphics cards are developing rapidly,and deep convolutional neural networks are excellent in many problems such as target recognition,speech recognition,and natural language processing.This paper proposes to classify high-speed congested pictures based on deep convolutional neural networks.First of all,according to the congestion classification criteria of the "Urban Traffic Management Evaluation Index System",combined with the actual situation,the highway pictures are divided into three levels: congestion,saturation,and unobstructed.The number of congestion samples is 10689,the saturated sample is 9645,and the unobstructed sample There are 10177.Two classic CNN models,Alex Net and VGG16,were used to evaluate their classification accuracy in this data set.The VGG16 model has the highest accuracy on the test set,reaching 91.04%,which is 3.81% higher than Alex Net.Then,this paper improves the VGG16 network from two perspectives of transfer learning and network structure.The first is to pre-train the model through Image Net and fine-tune its applied data set,which improves the overall test accuracy of the original model by 4.74%.Then,on the basis of it,the migration method using convolutional network for feature extraction and support vector machine(SVM)for classification is verified.Experimental results show that the method based on feature migration can improve the calculation speed during the testing process,and the classification accuracy has not changed significantly.Finally,this paper establishes an algorithm based on the improved VGG16 network model.The improved VGG16 network model adds a bottleneck layer and the SE Block to reduce the number of parameters and reduce the computational complexity,while expanding the effective weight response and improving the classification accuracy.rate.Using Image Net pre-training parameters and fine-tuning using the data set in this paper,the final test accuracy is as high as 98.04%,which is 2.26% higher than that of using only fine-tuning VGG16 method.The algorithm proposed in this paper can be effectively used for the traffic state recognition of highway images.The algorithm has high accuracy and fast recognition speed.It can be applied to practice and reduce the loss of human resources and materials caused by congestion.