Research On Traveler Transportation Mode Identification And Destination Prediction Based On GPS Data

With the popularity of smart phones and other portable GPS devices,it is possible to efficiently collect traveler GPS data containing rich temporal and spatial information.On the one hand,GPS trajectory mining can help government understand the travel mode of urban travelers and provide support for policies such as urban traffic scheduling and urban planning;On the other hand,it can also help draw traveler portraits and serve tasks such as locationbased advertising recommendation and traveler destination prediction.This thesis focuses on three key tasks in GPS trajectory mining,traveler trajectory segmentation,transportation mode identification and destination prediction.The main contributions are as follows:(1)We have proposed a spatio-temporal features integrated trajectory segmentation framework(Spatial-Temporal Change Points Detect).The model extracts the spatial morphological features and temporal motion features of GPS readings,identifies candidate segmentation points by using multi-view deep network,and then detect the segmentation points through mean filtering and mean shift clustering.The F1-score of trajectory segmentation of STCPD is increased compared with the baseline methods.(2)We have proposed a geographic information integrated semi-supervised transportation mode identification model Geo SDVA(Geo-information Semi-supervised Dirichlet AutoEncoder).The model has combined the trajectory motion characteristics with the surrounding geographic information to form the feature of trajectory,and utilized the semisupervised model based on Dirichlet Variational Auto-encoder to identify the transportation mode using a large number of unlabeled trajectories and a small number of labeled trajectories.The identification accuracy of Geo SDVA is improved compared with the baseline methods on two real GPS data sets.(3)We have proposed a regional semantic and transportation mode integrated destination prediction mode SEDP(Semantic Embedding Destination Prediction).The model utilizes pre-training model to extract semantic features from traveler trajectories,and predict destination with historical popular destinations through Bi-GRU model.The prediction MHD error of SEDP is reduced compared with the baseline methods on two real GPS data sets through experimental analysis.To sum up,focusing on the key tasks in trajectory mining,we have proposed STCPD model,which improved the segmentation performance of trajectory segmentation based on traveler behavior by integrating temporal and spatial feature.We have proposed Geo SDVA model,which solved the problem of transportation mode identification problem with few labeled trajectories through variational auto-encoder.And a destination prediction model SEDP has been proposed to predict traveler destinations more accurately by combining semantic features.This research can be applied to the fields of transportation planning and tourism management,and has practical significance.