Neuromorphic-classical hybrid framework for real-time augmented reality processing: adaptive stream fusion with dynamic threshold modulation and uncertainty-aware sensor integration

Abstract

Contemporary augmented reality systems must reconcile three competing demands: latency low enough to remain imperceptible; visual fidelity high enough to convince a viewer that virtual content belongs in the physical scene; and power draw low enough to enable extended mobile use. This study narrows its focus to three specific bottlenecks that recur across conventional pipelines and proposes an algorithmic remedy for each, evaluated through implementation and controlled synthetic execution rather than field deployment. The first bottleneck is feature extraction, which conventional pipelines were found to spend 42% of total processing time on. The second is unreliable multimodal sensor fusion, whose spatial-tracking accuracy, as the same profiling found, could degrade by 68% when lighting or motion departs from laboratory conditions. The third is reactive content delivery, whose 180 ms loading delays break immersion each time a user turns their head. The design targets that motivated these numbers—3D molecular visualization at 120 ± 3.2 Hz for education, anatomical overlays precise to 1.1 ± 0.15 mm for remote surgical guidance, assembly-instruction positioning within 1.8 ± 0.2 mm for industrial maintenance, and 28 ± 4 ms contextual content delivery for location-based entertainment—are presented strictly as potential applications suggested by this algorithmic evaluation, not as demonstrated outcomes. Validating the framework on commercial hardware such as Xreal One Pro, Apple Vision Pro, and Meta Quest Pro remains a planned next step beyond the present algorithm-level, synthetic-dataset scope.