par Machado Garcez Duarte, Mariana 
Président du jury Mühlberg, Jan Tobias
Promoteur Sakr, Mahmoud
Co-Promoteur Zimanyi, Esteban
Publication Non publié, 2026-04-23
Président du jury Mühlberg, Jan Tobias
Promoteur Sakr, Mahmoud
Co-Promoteur Zimanyi, Esteban
Publication Non publié, 2026-04-23
Thèse de doctorat
| Résumé : | The propagation of Internet-of-Things sensors aboard moving objects has led to continuous spatiotemporal data streams that demand on-device and low-latency analysis. The underlying systems for processing this streaming data, however, are ill-prepared. On the one hand, standard stream processing engines lack support for spatiotemporal operations. On the other hand, existing libraries for spatiotemporal data are optimized for historical data rather than real-time processing. Spatiotemporal streaming workloads are intrinsically more demanding than relational event processing. Spatial predicates require geometry-specific algorithms rather than scalar comparisons, while spatiotemporal reasoning adds temporal constraints to these computations. Moreover, trajectory construction requires ordering, buffering, and interpolation under tight latency budgets, and motion features (e.g., speed, heading, stops) require sub-trajectory processing, which increases CPU, memory, and synchronization costs.To address these challenges, two systems were implemented: MobilityFlink and MobilityNebula. MobilityFlink, an extension of Flink for spatiotemporal stream processing in the cloud, served as the initial prototype to validate spatiotemporal stream operators within a well-established cloud engine. Building on these insights, MobilityNebula was designed by migrating to NebulaStream to leverage distributed edge-to-cloud execution and support continuous, low-latency processing in constrained, intermittently connected deployments. MobilityNebula is a mobility stream processing system designed for the edge-to-cloud continuum. We introduce a window-aware trajectory type system, a lift-combine-lower aggregation pattern for trajectory construction at stream rate, and a lightweight operator integration that enables Moving Objects Database trajectory semantics to operate on constrained edge devices. MobilityNebula was evaluated by deploying the system on fog and edge devices and ingesting data from the Belgian Railway Operator/Société Nationale des Chemins de fer Belges trains, executing real-time mobility queries such as brake system monitoring and high-risk zone proximity monitoring directly at the edge. In our evaluation, MobilityNebula sustained near real-time latencies for queries while ingesting a 20k events/s stream on an edge device such as a Raspberry Pi 5.Additionally, we survey outlier detection as it is an essential step in data preprocessing that ensures the integrity and validity of data analyses. We focus on outlier points within individual trajectories, i.e., points that deviate significantly within a single trajectory. We experiment with ten open-source libraries to comprehensively evaluate available tools, comparing their efficiency and accuracy in identifying and cleaning outliers. The experiment considers the libraries as they are offered to end users, with real-world applicability. We compare existing outlier detection libraries and introduce a method for establishing ground-truth. Furthermore, we use our findings to implement Kalman filter into MobilityNebula |
