A Comprehensive Survey of Deep Learning–Based Object Tracking for Augmented Reality in Complex Real-World Scenes

Augmented Reality (AR) systems critically depend on accurate, temporally stable, and computationally efficient object tracking to maintain geometric alignment and perceptual coherence between virtual content and the physical world. As AR technologies transition from controlled laboratory prototypes to large-scale deployment in industrial, medical, and consumer scenarios, tracking must operate robustly in complex real-world environments characterized by dynamic objects, occlusions, illumination changes, fast motion, and strict computational constraints. Traditional geometry-driven tracking pipelines often degrade under such conditions, motivating increased adoption of deep learning based approaches. This survey provides a comprehensive review of deep learning based object tracking for AR in complex real-world scenes, with particular emphasis on system-level considerations. Object tracking is treated as a central perception primitive that underpins stable AR experiences and interacts tightly with modules such as visual simultaneous localization and mapping, depth and geometry estimation, and semantic scene understanding. In contrast to prior surveys that emphasize algorithmic accuracy in isolation, we explicitly analyze AR-specific constraints including real-time latency, temporal stability, energy efficiency, long-term robustness, and deployment on mobile and wearable platforms. We review major tracking paradigms, representative datasets and benchmarks, AR-centric evaluation criteria, and common failure modes observed in practice, and we outline future research directions toward scalable, reliable, and trustworthy AR tracking systems.