Research On Key Issues Of Individual Intelligent Travel

With the continuous development of mobile internet and artificial intelligence,people are pursuing more comfortable and more intelligent daily travel patterns.With the increase of the types of transportation modes in modern society,the complexity of traffic system becomes increasingly complex.Therefore,users’ efficient and intelligent travel has become a growing concern.However,due to the complex urban road network,the nonlinear change of traffic flow dynamics,and the influence of climate,emergencies and other factors,people’s efficient and comfortable travel experience is severely affected.Therefore,the research on the key technologies of individual intelligent travel has important theoretical significance and practical application values.Although many studies on intelligent traffic problems have already existed,there are other problems such as insufficient mining of spatio-temporal information,low accuracy of traffic pattern recognition and low accuracy of recommendation of points of interest in road traffic prediction.Based on comprehensive understanding of transport properties of individual travel data and information,this article conducts spatio-temporal data traffic flow forecasting and traffic speed prediction,traffic modes recognition based on multi-source sensor information,and user interest recommending based on cross modal.The main contribution of this thesis are as follows:(1)Because of the problems of insufficient dynamic correlation information of each sensor node and remote sensor information mined in current traffic speed prediction,we proposed a spatio-temporal based on a dynamic multi-graph convolution cyclic network.DMGCRN algorithm can obtain the similar spatio-temporal patterns between the long-distance nodes in the traffic network by mining the potential relations among the long-distance nodes with similar spatial structure.The problem of insufficient expression of the dependence between long-term time dimension and global spatial dimension in existing traffic flow prediction.We propose graph convolution long short-term memory continuous spatiotemporal sequence representation learning algorithm based on enhanced memory attention.The attention mechanism and graph convolution longterm and short-term memory used to mine the correlation between the trend of time change and the global and local space.The proposed algorithms verify on open data sets.The prediction results are superior to the baseline experiments.(2)Because of the problems of inadequate inter-temporal completeness feature mining and semantic characteristics understanding for multimodal fine-grained traffic mode recognition,we proposed the short-long time domain features deep mining algorithm.By obtaining the short time and long-time features from four types of sensors data,it effectively overcomes the problem of incomplete feature mining.It is difficult to achieve high-precision traffic mode classification due to insufficient exploration of the influence of data noise and correlation between features.To solve the problem we propose the multi-scale spectral feature mining from frequency domain perspective,which use the network-in-network and dilated convolution to obtain multi-scale feature representation.Under the condition of resource limitation,a small amount of sensor data is used to transform the one-dimensional data into two-dimensional spectrum by wavelet transform,which further enhances the signal expression of a small amount of data and reduces the influence of data noise.Experimental results show that the proposed transportation mode recognition algorithms can achieve high-precision traffic pattern recognition in different resource environments.(3)The problem is insufficient mining of spatio-temporal relations containing complex location information and check-in sequence data.We propose a multi-modal point of interest recommendation algorithm(AutoMTN)enhanced by an autocorrelation mechanism transformer network.The AutoMTN algorithm builds a two-channel self-correlation model based on the location information of interest points and the category information of interest points in the trajectory of individual user to capture the correlation between the neutron sequence of individual user trajectory.Cross-modal self-correlation algorithm is used to mine the interaction between the position sequence of interest points and the category sequence of interest points of individual users.Experimental results verify the effectiveness of the multi-modal point of interest recommendation algorithm based on autocorrelation enhancement.Individual intelligent travel is also an important research content of mobile computing and ubiquitous computing.During user’s intelligent travel,problems such as road flow prediction of individual travel,transportation mode recognition,and recommendation of individual travel destinations are encountered.The effective solution of the above three key problems play an important role in solving the intelligent travel problem of users.In this paper,we use large-scale public data sets or self-built data sets to verify sufficient experiments which are carried out based on the algorithms proposed above.Meanwhile,comparative analysis and evaluation are carried out with advanced baselines to verify the validity of each algorithm.