Project R-16336

Robust Context-aware object detection for edge devices (Research)

Object detection has achieved remarkable performance on curated benchmarks; however, detectors deployed in real-world edge environments remain brittle under challenging conditions such as occlusion, small objects, cluttered backgrounds, and distribution shifts. These issues are particularly critical in embedded vision systems where computational constraints enforce low input resolutions, compact models, and strict latency budgets. A fundamental limitation of current object detection systems is their frame-centric perception paradigm. Most detectors rely primarily on local visual evidence within individual frames, while ignoring broader contextual signals that could disambiguate uncertain observations. In real-world environments, however, object interpretation is inherently contextual: spatial layout constrains plausible object locations, temporal continuity stabilizes perception under occlusion, and semantic relationships between objects provide additional cues for interpretation. Although recent research has explored contextual reasoning in computer vision, existing approaches typically rely on computationally expensive architectures such as transformer-based models or graph reasoning networks, which are often impractical for deployment on resource-constrained edge devices. Consequently, there is currently no principled framework that formalizes context and integrates it efficiently within real-time object detection systems for edge AI environments. This gap motivates the central research question of this work: How can spatial, temporal, and semantic context be formally represented and exploited to improve object detection robustness while maintaining real-time performance on edge devices?

01 January 2026 - 31 December 2029