摘要
Search-and-rescue (SAR) operations require UAV systems that can flexibly search for mission-relevant targets and deliver actionable information under challenging aerial imaging and communication conditions. However, high-altitude UAV imagery often contains small, partially occluded, or visually ambiguous targets, while SAR missions may involve object categories that cannot be fully predefined before deployment. This paper presents a promptable open-vocabulary UAV event perception framework for SAR. Operators specify target concepts using short natural-language prompts, and the onboard system detects the queried targets, maintains their identities, estimates their geo-locations, and consolidates repeated observations into persistent event records. To support bandwidth-constrained operation, compact event metadata and target-centered evidence are transmitted before broader visual context. The perception module adapts promptable concept segmentation to aerial SAR imagery, improving small and irregular target localization while preserving prompt-based target specification beyond the adapted categories. Experiments on public aerial datasets and self-collected day--night fire-source data demonstrate improved promptable perception for small and irregular SAR targets, while outdoor field tests with the integrated UAV platform validate onboard event generation, geo-localized reporting, and progressive evidence delivery under a constrained 4G uplink.