A Proposed Framework for Robust Visual Product Identification in Dynamic Environments

Abstract

Reliable visual product identification remains a challenging task in dynamic indoor environments, where variability in viewpoint, illumination, and object placement limits the applicability of conventional scanning-based approaches. This is particularly relevant in high-mix low-volume intralogistics settings, where incomplete observations and inconsistent acquisition conditions are common. This paper proposes a modular framework for robust visual product identification based on machine-readable codes. The approach formulates identification as a system-level reasoning problem, integrating detection, geometric normalization, decoding, validation, recovery, and temporal aggregation within a unified architecture. Frame-level observations are treated as uncertain evidence, which is accumulated and constrained at the sequence level in order to support reliable identification decisions. The individual components draw on established computer vision and decoding techniques; the contribution lies in their system-level integration with explicit uncertainty handling and temporal reasoning, rather than in new detection, decoding, or fusion algorithms. The proposed framework emphasizes robustness through structured integration of visual evidence rather than isolated recognition performance, providing a foundation for perception driven identification in visually dynamic environments. A preliminary evaluation on a controlled test set, shows that detection performance is close to ceiling, while frame-level decoding alone succeeds on only about one third of ground-truth instances (35.6%); aggregating evidence across frames within the proposed sequence-level reasoning layer raises identification success to 73.5%, providing initial evidence that the proposed architecture improves reliability over frame-level decoding alone.