UCLA

Actions and decisions in physical environments often depend on detecting patterns of interest, known as complex events. In distributed sensing systems, complex event detection may require combining unstructured sensor data with structured measurements and inferences produced by sensing and recognition models. Complex events may unfold from seconds to weeks and across spatial scales ranging from a single room to an entire city.Existing systems commonly assume that relevant operations, output representations, and execution plans have been selected in advance. However, observations may support competing explanations, sensor streams may expose more information than an application requires, and needed evidence and computation may change as an event progresses. This dissertation argues that distributed sensing systems should use the event and its application context to revise which events remain plausible, determine what information may be disclosed, and adapt sensing and computation as the event unfolds.First, we address the challenge of discovering and localizing complex events when distributed observations support competing event explanations. Our system, IncidentLens, performs city-scale emergency monitoring by maintaining hypotheses about an incident and relationships among its observed effects. Evaluation on urban and synthetic data shows that incident priors and hypothesis management support discovery from sparse and partial observations.Second, we address the challenge of balancing utility with privacy and data-use constraints for complex event applications in smart environments. Our system, privMediator, composes sensor-processing operations, verifies that outputs support the application task and applicable information flow rules, and selects a feasible representation that avoids unnecessary disclosure. Across smart space tasks, its outputs avoid configured policy violations and are judged more appropriate than raw media and developer-authored alternatives.Third, we investigate how complex event sensing and computation should adapt as events progress and system conditions change in tactical environments. Through FABLE, a Python research prototype and controlled MQTT/replay testbed, we explore whether progress through a developer-authored complex-event workflow can determine which predefined sensing and inference work is useful next. Experiments on 89 recorded traces spanning nine separately authored workflows examine the recognition–computation tradeoff of progress-driven activation within these curated configurations, including a controlled network-disconnection case.Together, these contributions examine complementary decisions in complex event detection for dynamic distributed environments, such as smart cities: distinguishing among competing event explanations, disclosing information appropriate to the application context, and using event progress to guide sensing and computation.