Adaptive Depth-Aware Real-Time Object Perception for Contextual Mobile Augmented Reality

Abstract

Real-time object detection in mobile augmented reality (AR) can produce visually cluttered overlays when all detected objects are presented with equal visual priority. This paper presents Adaptive Depth-Aware Confidence Reweighting (ADCR), a lightweight detection-level mechanism designed to prioritize spatially relevant objects and suppress visually distracting overlays in mobile AR. ADCR aims to reduce visual clutter, improve perceptual clarity, and support more focused and efficient interaction by ensuring that the most relevant detections receive greater visual priority. Class-aware, one-to-one intersection-over-union matching establishes correspondence between consecutive processed frames, while active-cue renormalization maintains consistent scoring when depth or temporal evidence is unavailable. Unlike binary distance filtering, ADCR continuously modulates confidence without modifying the underlying detector. The framework integrates YOLOv8n, Unity Sentis, AR Foundation, and ARCore within a fully on-device mobile AR pipeline. Evaluation included implementation-matched baselines, quantitative ablation, hyperparameter and depth-perturbation analyses, viewpoint-dependent object prioritization, and a within-participant study with 34 participants. ADCR produced substantially greater clutter suppression than fixed-depth filtering, and component analyses confirmed complementary contributions of depth, scale, and temporal association. The application maintained approximately 28–30 rendering FPS on the Samsung Galaxy S22 Ultra. Holm-corrected paired analyses confirmed significant improvements in task efficiency and user experience, demonstrating ADCR’s practical value for clearer, spatially prioritized mobile AR interaction.