Towards urban mobility-based activity knowledge discovery: interpreting motion trajectories

Abstract

Understanding travel behaviour is important for an effective urban planning and to enable different context-aware mobility service provisions. To this end, it is essential to model different mobility-based activities in available trajectory data. However, the semantics of activity varies from context to context, which poses a challenge for developing a connected knowledge flow for different services.

Currently, such mobility-based information is typically collected through manual paper-based surveys. These surveys preserve context, but come with their own inherent quality issues, and are expensive in comparison to data analytics methods. To address this issue this research leverages the emerging concept of smartphone-based travel surveys that collect people’s movement behaviour in terms of raw trajectories.

This research proposes an ontological framework that can model activities in a hierarchical manner adapting to different contexts and thereby addressing the challenges of trajectory data analytics mentioned above. This research also explores how raw trajectories collected by a smartphone can be interpreted to generate mobility information (e.g., transport modes, trips). While interpreting the trajectories this thesis models uncertainties that may exist during people’s travel behaviour and interpretation process.

In this research, a particular focus is given to knowledge representation, that is understanding urban movement behaviour from detecting transport modes in trajectories. One presented form of knowledge representation is a fuzzy logic based approach to mode detection. The knowledge representation is essential to extract semantics related to a given activity. This research also introduces the concept of near-real time mode detection and investigates the performance of a purely knowledge-driven model works effectively in a near-real time scenario. Since a knowledge-driven model at different temporal granularities while detecting a given transport mode. The knowledge-driven model that works in offline, typically requires kinematic features computed over sufficiently long segments. But in near-real time these segments must be shorter and requires the model to be adaptive. To address this issue a machine learning based model has been deployed, which can learn from the historical data, and work in varied conditions. But machine learning models work as a black box and cannot explain their reasoning scheme owing to a semantic gap in the activity knowledge base. On the other hand, a fuzzy logic based model can explain its reasoning scheme but cannot adapt to varying conditions. To bridge the trade-off between these approaches this research proposes a hybrid knowledge-driven framework that is capable of self-adaptation and explaining its reasoning scheme. The results show the hybrid model performs better than a purely knowledge-driven model and works at par with the machine learning models for transport mode detection. This research also justifies a hybrid approach can model the activity in a consistent and adaptive manner while explaining the semantics related to different mobility-based activities.

In this research different uncertainties related to a motion trajectory interpretation process have been addressed. A particular focus is given on modelling the temporal uncertainties that exist between predicted, scheduled and reported trips. Such a temporal uncertainty quantification measures the reliability (or uncertainty) in an inference process in the interest of information retrieval at different contexts. Considering the lack of semantics in GPS trajectories an investigation is also made whether incorporating low sampled IMU information in addition to a GPS trajectory can improve the accuracy. This research also identifies existing trajectory segmentation approaches (e.g., clustering-based or walking-based approaches) are subjective and thus lacks adaptivity. In order to address these issues a novel state-based bottom-up trajectory interpretation model is developed, which can generate mobility information at different temporal granularities. The model also demonstrates its efficacy, flexibility, and adaptivity over the existing top-down approaches This research also demonstrates that using a GPS trajectory, it is possible to generate modal state information comparatively at a coarser granularity but shorter than the time required to generate information from a historical GPS trajectory. The response time is subject to a particular application domain.

The research presented in this thesis has a potential to improve the background intelligence in smartphone-based travel surveys and smartphone-based travel applications facilitating mobility-based context-aware service provisions where the notion of activity is prevalent at different granularities. However, this research cannot distinguish composite activities, which require future work. With the emergence of Web 2.0 and ubiquitous location sensing technologies, the location information can come from various sources with the different level of inaccuracies and space-time granularities. The models developed in this research currently work best on GPS trajectories sampled at 1 Hz to 2 Hz frequency, which may be enriched with IMU information. However, the models need some adjustments and incorporations of additional features and rules when the location information comes not only from GPS but also from GSM, Wi-Fi, smart-card. The models developed in this research are flexible, transparent and offer provisions for further enrichment of raw trajectories and extract finer activity information. This research has a potential to understand mobility patterns at an aggregate and a disaggregate level, and thereby serve different application domains e.g., personalized activity recommendations during a travel, emergency service provisions, real-time traffic management and long term urban policy making.