We present VIGIA-E(Vision-based Inference for Guided Interest Attention on Edge devices), a lightweight object detection strategy for online surveillance in wide-area scenes using resource-constrained edge hardware. Designed for typical surveillance settings, such as public spaces or coastal infrastructures monitored by drones or elevated cameras, it addresses the challenge of detecting small, distant objects caused by perspective and scale. The method adopts a two-stage pipeline: a global low-resolution pass followed by selective high-resolution inference over dense regions identified via a grid-based estimator. This strategy reduces redundant computation while preserving detection accuracy, enabling efficient deployment on embedded platforms. Unlike conventional multi-inference approaches, VIGIA-E brings online feasibility to density-driven small object detection on edge systems, a crucial requirement for modern smart surveillance. We evaluate VIGIA-E on two complementary datasets: the domain-specific Anfi dataset and the VisDrone benchmark. In both cases, it achieves favorable trade-offs between accuracy and computational cost compared to representative multi-inference baselines. Additionally, VIGIA-E has been deployed in a real-world coastal surveillance system, demonstrating operational viability. While representing an initial iteration of our framework, it establishes a solid foundation for more advanced systems under development.