Study On Traffic State Recognition Of Urban Arterial Road From High Altitude Videos Based On Deep Learning

With the continuous development of China’s economy and the continuous growth of China’s car ownership,the current problem of road traffic congestion has become increasingly prominent.As the main artery of urban traffic system,the main road is the primary focus of urban traffic congestion control.Urban arterial roads cover a wide range and have a large impact.Conventional induction coils and video monitoring equipment are not conducive to the overall observation of the changes in the traffic status of the arterial roads.In recent years,the gradual popularity of high-altitude video(including UAV video,high-altitude monitoring video,Hawkeye video,etc.)provides a new perspective for the traffic research of the arterial roads.At present,there are many achievements in the field of video target detection and tracking,but there are relatively few targeted researches on high-altitude video traffic state recognition,and the methods used are mainly based on the traditional image feature detection.With the rapid development of artificial intelligence,deep learning theory and method provide a new solution for many fields.Therefore,it has become a hot trend to introduce deep learning theory into the field of high-altitude traffic video,research and solve traffic problems.To sum up,this paper attempts to use the self encoder model,three-dimensional convolution neural network model and deep neural network model in the deep learning method to solve the problem of urban arterial traffic status recognition and prediction based on high-altitude video,to provide new technical support for arterial traffic status monitoring and control,and to provide new perspectives and methods for easing traffic congestion.The main contents of this paper are as follows:(1)This paper classifies and summarizes the current mainstream deep learning theory,and analyzes the classic neural network model of these algorithms,which lays a theoretical foundation for the construction of deep learning model in this paper.(2)For image traffic state recognition,this paper optimizes model structure parameters from three aspects: input data dimension,number of hidden layers and dimension reduction data dimension,strengthens the extraction of image features by self encoder,and realizes efficient compression of massive image features.Finally,based on the dimension reduction data and K-means clustering algorithm,the traffic state is effectively identified.(3)Aiming at video traffic state recognition,this paper introduces C3 d,a threedimensional convolutional neural network model commonly used in the field of behavior recognition,into the research of video traffic state recognition,and regards the traffic state as a special behavior action for video recognition.Therefore,the video traffic state recognition data set is constructed and optimized based on the three-dimensional convolution neural network C3 d expansion model,including adjusting the convolution layer structure,optimizing the plane convolution size and optimizing the video convolution core depth.Finally,the optimal model 3dcnn * is tested and verified.(4)Aiming at video traffic prediction,this paper proposes a traffic state prediction method based on 3dcnn and DNN,which transforms the limited traffic state prediction problem into classification problem,and realizes short-term video traffic state prediction.Through the research of the above content,this paper draws the following conclusions:(1)In the aspect of image traffic state recognition,the AEK model based on the combination of self-made device and K-means in this paper not only has a good ability of image traffic state recognition,but also can reduce model parameters,save labor costs and greatly improve the efficiency of model detection.(2)In the aspect of video traffic state recognition,3dcnn * is better than the common3 D CNN model and 2D convolution network model.It shows that the traffic state recognition model constructed in this paper has good performance in two-dimensional convolution and three-dimensional convolution,and has high generalization ability.(3)In the aspect of video traffic prediction,the traffic state prediction model based on 3D convolution neural network depth neural network(3dcnn-dnn)can accurately predict the video traffic state,which provides a new method for short-term data traffic state prediction.