Research On Trajectory Prediction And Intention Mining

The increasing prevalence of positioning devices and video capturing equipments has made it possible to collect a vast amount of temporal and spatial movement data. For example, many people like to share their locations with their friends, which generates many check-in data. As another example, vehicles are photographed when they pass the surveillance cameras on the roads and structured vehicle passage records (VPRs) are subsequently extracted from the pictures using OCR (Optical Character Recognition). Both check-in data and VPRs are typical temporal and spatial trajectory data, and they both consist of three attributes:object (e.g., human, vehicles), location and time-stamp. These data contain huge values and produce fruitful results in many fields, e.g., urban computing, route planning and location prediction.In this thesis, we make an indepth research on the related problems in the field of temporal and spatial trajectory mining. First, we focus on the next location prediction problem. In practical applications, once knowing the next locations in advance, we can recommend more reasonable driving routes for users and recommend some advertising information in the predicted area. To predict next locations, we first propose the Global Markov Model (GMM) and the Personal Markov Model (PMM). GMM uses all available trajectories to discover the collective behaviors; PMM focuses on mining the individual patterns of each moving object. The two models are integrated with linear regression to generate the final predictor (NLPMM). In addition, time factor may affect human mobility patterns, and we then seek to further improve the accuracy of prediction by considering the time factor, with a focus on clustering the trajectories in different time periods, and present three methods to train the time-aware models to predict next locations.Furthermore, we propose a novel Markov model (objectTra-MM) exploiting the similarities between objects and the ones between trajectories to predict next locations. ObjectTra-MM consists of two models:object-clustered Markov model (object-MM) and trajectory-clustered Markov model (tra-MM). object-MM first clusters similar objects based on their spatial localities, and then builds variable order Markov models with the trajectories of objects in the same cluster. Tra-MM considers the similarities between trajectories, and clusters the trajectories to form the training set used in building the Markov models.Finally, we mine the intentions underlying such trajectory data in order to better understand the mobility patterns of people. In this paper, we propose a novel probabilistic model called the Intention of Movement (IoM) to model the generative process of trajectories. It is based on three key observations:(1) intentions of movements are reflected by the sequences of locations in the trajectories; (2) different people have different intentions; and (3) trajectories tend to be cyclical and change over time. Thus, IoM is built in such a way that can help discover the intentions underlying a trajectory by jointly modeling these aspects, i.e., location sequences, objects and time. For completeness, we also examine some alternative methods that consider only a subset of these aspects. We perform thorough empirical studies on two real datasets, and the experimental results confirm the effectiveness of the proposed models.