Research On Trajectory Destination Prediction Algorithms

With the development of communication technology and the widespread application of mobile GPS devices,a large amount of trajectory data has been generated,and these data contain rich behavioral patterns of moving objects.Predicting destinations with high accuracy and efficiency historical trajectories can have a wide range of applications in multiple fields and scopes,and has become a current research hotspot.This article studies how to better perform destination prediction,mainly including the following aspectsFirst,In order to solve several problems of the existing trajectory prediction methods:1)the trajectory is modeled as one-dimensional sequence and the two-dimensional spatial relationship of the trajectory points is ignored.2)In dealing with trajectories at a single scale,the multiscale nature of trajectories is ignored.3)Every part of the track is of the same importance,but not all parts of the track are meaningful.we propose two destination prediction models T-CONV-Basic and T-CONV-LE.The T-CONV-Basic model models the trajectory as a two-dimensional image,and uses a convolutional neural network to extract the trajectory image multi-scale information.Further,we analyzed that the trajectories near the starting point and the current point by gradient visualization to contribute the most to the result,so we propose a trajectory destination prediction model based on local enhanced convolution(T-CONV-LE),T-CONV-LE can more effectively extract the features of the local area of the trajectory and avoid the problem of data sparsity.Experiments on real datasets show that T-CONV-LE can achieve better results than T-CONV-Basic and other methodsSecondly,in order to solve the local area size selection problem of the T-CONV-LE model,a locally enhanced convolutional neural network destination prediction model(T-CONV-LE-MUL)based on the attention mechanism is proposed to extract the trajectory of the local area Flow information,and use the attention mechanism to take the trajectory flow information as input to realize the adaptive selection of the size of the area,so as to achieve the fusion of the trajectory image information of different areasThen,we propose an algorithm for extracting road network data containing road topology information using a large number of vehicle trajectory points.This method uses denoising,sampling,interpolation,and morphological processing at multiple resolutions through trajectory point data.,Efficiently extract road network information.After obtaining the road network data,the convolutional neural network is used to extract the topological features of the road network data,and the road network features are integrated into the trajectory destination prediction algorithm.A vehicle trajectory destination prediction algorithm(T-CONV-LE-MUL-ROAD)to improve the accuracy of trajectory destination predictionFinally,based on the theoretical research,a large number of experiments were performed using the Proto data set and the T-drive data set to verify the T-CONV-Basic,T-CONV-LE,T-CONV-LE-MUL T-CONV-LE-MUL-ROAD model effect.Compared with the existing methods,experiments are performed on accuracy,parameter sensitivity,and running time.The experiments prove that these algorithms have achieved good results.