A Study on map-matching and map inference problems

Abstract

The recent upsurge of GPS(Global Positioning System)-equipped devices has enabled the tracking of users'/vehicles' locations. Meanwhile, the emergence of various location-based services emphasises the crucial role of the underlying digital map. However, the intrinsic inaccuracy of both maps and GPS trajectories has been a major issue that hinders the development of location-based applications, like navigation, location-based recommendation and traffic analysis. Several techniques are proposed to deal with the data quality issues in maps and GPS trajectories, respectively: (1) To avoid the excessive cost of ground surveying for map construction, the map inference algorithm aims to construct a digital map from GPS trajectories automatically. (2) To ensure the recency of digital maps, the map update algorithm focusses on updating an existing map using recent GPS trajectories. (3) To get rid of the GPS trajectory errors, the map-matching algorithm aligns a trajectory to the underlying map to find the user's actual travel path. In our thesis, we mainly focus on surveying the above techniques as well as proposing new solutions for solving the aforementioned data quality issues. Overall, our contributions are listed below.1. We conduct a comprehensive survey and experimental study of existing map inference algorithms. Due to the labour intensity of traditional map creation and the frequent road changes nowadays, map inference is deemed to be a promising solution to automatic map construction and updates. However, existing map inference algorithms suffer from low GPS accuracy, which makes the quality of the constructed map unsatisfactory. In this thesis, different from previous surveys, we (1) include the most recent solutions and propose a new categorisation of methods; (2) we study how different types of GPS errors affect the quality of inference results; (3) we compare the existing map inference quality measures on both real dataset and synthetic datasets, which are generated by our proposed data generators, regarding their ability to identify map quality issues. Overall, our study provides a guideline about (1) which inference method should be considered for each type of applications, (2) what trajectory quality should be guaranteed for map inference and (3) the direction of future work for quantitative map quality measures.2. We review the existing map-matching algorithms and study their performance under different data quality scenarios. As an indispensable pre-processing step for trajectories, the trajectory map-matching problem has been an ongoing research topic for more than two decades. In this thesis, we summarise and classify the existing map-matching algorithms based on their map-matching models, working scenarios and input data features. In addition, we conduct extensive experiments on several representative map-matching algorithms to compare their performance under various working scenarios, data settings and performance requirements. The experiments are done on both real and synthetic datasets with different scales to (1) characterise the strengths and weaknesses of each algorithm category, (2) reveal how data quality issues affect the map-matching performance and (3) identify the remaining challenges in map-matching problems.3. We propose a co-optimisation framework that aims to solve the data quality of both GPS trajectories and maps simultaneously. Despite that many map-matching and map inference/update techniques have been proposed to deal with data quality issues on GPS trajectories and maps, respectively, calibrating one of the datasets relies on the other as a reference. Those reference data are required to be accurate, which is unobtainable in practice. Therefore, we propose a map-trajectory co-optimisation framework that takes the inaccurate map and trajectory data as inputs and mutually increases the accuracy of both. In our framework, both map-matching and map updates are run iteratively, and we propose two scores for each new map update, namely influence score and confidence score, to ensure the map is updated correctly and in a consistent way. Besides, our framework accepts most of the existing map-matching and map inference algorithms as candidate solutions in matching and update phases and receives straightforward performance boost through our framework.4. We develop a map service platform that supports the aforementioned data preprocessing/cleaning procedures. Considering the close relationship between map-matching and map inference/update and their broad applications, there is a lack of open-source tools providing those map-based data cleaning processes. In this thesis, we introduce the main features and functionalities of our map service platform, which supports multiple data cleaning processes, including map-matching, map inference and co-optimisation. Our platform provides various solutions for each type of cleaning process, which can be used for both future research comparisons and industrial applications.