HA-Pos: Hierarchical Prompt-Guided Adaptive Detection for Cross-view Visual Positioning System

With the rapid proliferation of Location-Based Services (LBS), achieving high-precision self-positioning on consumer-grade mobile devices—such as smartphones and civil drones—remains a critical challenge, particularly in GPS-denied or multipath-prone urban environments. This paper proposes HA-Pos, a novel hierarchical adaptive prompting mechanism enhancing the Cross-view Visual Positioning System (CVPS) for consumer electronics. The proposed method enables target specification via a user-defined click on a query image captured by a consumer terminal, subsequently locating that object within corresponding satellite reference imagery. Unlike traditional methods struggling with cross-view geometric distortions, HA-Pos incorporates a Hierarchical Prompt Query Encoder (HPQE). This encoder provides precise spatial guidance across various depth stages, significantly bolstering the ability to distinguish target objects from distractors. Building upon this, a Geometric Adaptive Decoupled Head (GAD-Head) is designed to improve geometric adaptability and positioning accuracy. The GAD-Head integrates deformable convolutions as a Deformation-Aware Module (DAM) to effectively capture geometric variations while independently optimizing regression and classification tasks. Extensive experiments demonstrate that HA-Pos achieves state-of-the-art performance on the CVOGL benchmark dataset.