Research On Crowd Traffic Prediction Method Based On Deep Learning

The accurate prediction of the crowd flow is of great significance for the government to effectively and reasonably managing the traffic resources,improve the travel experience of the masses,and improve the road traffic safety environment.However,because of the particularity of crowd flow data,it is a kind of complex spatiotemporal data,and the time dependence and space dependence must be taken into account when building the crowd flow prediction model.In addition,crowd flow is also affected by external factors such as events and weather.Therefore,it is not easy to accurately predicting the flow of people.The research object of traditional crowd flow prediction methods is usually the crowd flow pattern in a single region,but these methods only consider part of the spatiotemporal dependence of the crowd flow,but do not make a comprehensive and accurate measurement of the whole spatiotemporal dependence.Therefore,this paper will consider various factors affecting the change of crowd flow,and carry out modeling and prediction of crowd flow.In this paper,the urban area is divided into grid areas,and the crowd flow is represented and calculated by using the trajectory data of crowd flow,so as to obtain the tensor representation of the crowd flow on the time axis.To solve the problem of crowd flow prediction between grid areas,this paper proposes a corwd flow prediction model based on deep learning based on the densely connected concolutional neural network and the spatial-temporal prediction model(DCAST)of the gated cyclic neural network with attention mechanism,the model carries out reasonable unified modelling for various factors affecting the crowd flow,and also more accurately predicts the crowd inflow an outflow in the predicted area.In addition,in order to better capture the time dependence between regions,this paper designed a gated cyclic unit module with attention mechanism.The experimental results show that the gated cyclic unit module with attention mechanism plays an important role in the DCAST model proposed in this paper.DCAST model divides the time axis into three parts in order to solve the spatio-temporal dependence of crowd flow prediction in urban areas: Short-Term Dependence,Period Rule,the Long-Term Dependence,in the three modules respectively for regional population flow in the Short time dependencies,cycles,and long-term Dependence,each module by combining Dense Net and network structure,based on attention mechanism GRU helped the Dense Net network is used to describe the spatial dependencies,and based on attention mechanism GRU helped to describe time dependencies.Then,the output of the three parts is weighted and fused.In view of the external factors that affect crowd traffic,the DCAST model proposed in this paper uses a two-layer fully connected network to extract external features.Finally,the results of the fusion of the first three parts are integrated with the output of the external module,so as to predict the crowd flow in each region.In addition,this paper conducts experiments on DCAST model on two real data sets,namely Bike NYC and Taxi BJ of shared bike track in New York and Beijing,and the RMSE obtained is 5.53 and 15.70 respectively.Experiments show that the prediction results of DCAST model are more accurate and reliable than the traditional time series prediction method and other similar prediction methods based on deep learning.