Urban digital map construction and update for location based service in mobile sensing

Abstract

The inherent limitations in accuracy and diversity of urban digital maps are great obstacles to providing various location based services to mobile users. The emergence of mobile sensing, combining the power of both mobile computing and intelligent sensing, heralds a promising solution to the limitations on urban digital maps. In mobile sensing, urban digital maps are constructed and updated from multiple sensing data collected from mobile devices. The main disadvantage associated with this new technology is that it is difficult to extract accuracy and valuable information from a large amount of sensing data to construct and update urban digital maps. Therefore, it is crucial to design automatic indoor floor plan construction method, outdoor map update method, and urban safety index map construction method to support various location based services. In this thesis, we first propose the PlanSketcher system architecture to construct fine-grained and facility-labelled indoor floor plans with less energy consumption in smartphone. New landmark recognition approach is proposed to detect various landmarks. Then hallway topologies are constructed based on the sensing data, depth data and images through the proposed traverse-independent hallway construction algorithms. Because the indoor facility information is a crucial component of the indoor floor plan, we construct the room shape and label the recognized facilities in their corresponding positions to generate a complete indoor floor plan. To update the outdoor map, we propose an automatic store self-updating system through street views and sensing data crowdsourced from mobile users. A new weighted artificial neural network is developed to learn the underlying relationship between estimated positions and real positions to localize user's shooting positions. Then, we design a novel store name recognition method by considering two valuable features (i.e. the position of text in image and the colour histogram). In this way, we are able to recognize the complete store name instead of individual letters as previous study. Furthermore, we transfer the shooting position to the location of recognized stores in the map. To update changed stores in the map, we consider three updating categories (replacing, adding, and deleting) and estimate their positions based on the kernel density estimate model. Finally, to support urban safety related services, we propose an urban safety analysis system to infer safety index by leveraging multiple cross-domain urban location-based data. Effective spatially-related and temporally-related features are extracted from various data, including urban map, housing rent and density, population, positions of police stations, point of interests (POIs), crime event records, and taxi GPS trajectories. Then we present a novel feature fusion method to feed fused features into a spatial or temporal classifier, instead of treating features equally, leading to a high classification and inference accuracy. In addition, we design a new co-training-based learning method to accurately infer the safety index of each position in a city. The approach consists of two classifiers respectively modelling the spatial and temporal features which both influence the safety index.