The Lean Construction Visual Taxonomy (LCVT): bridging the semantic gap

The architecture, engineering, and construction (AEC) industry faces productivity stagnation due to ineffective production flow management. Although Lean Construction (LC) aims to minimize waste, manual monitoring lacks the high-frequency data required for timely control. Computer Vision (CV) offers automated monitoring but suffers from a "Semantic Gap," where models detect low-level objects but fail to interpret high-level Lean states (e.g., "waiting"). This study proposes the Lean Construction Visual Taxonomy (LCVT), a three-level hierarchical framework–Category, Indicator, Visual Definition grounded in Transformation-Flow-Value (TFV) theory. Crucially, the LCVT provides standardized class definitions to guide "zero-shot" prompt engineering in Vision-Language Models (VLMs). By injecting formal L3 definitions that address entity types, temporal thresholds (e.g., stationary >60 s), and spatial context into VLM models such as GPT-4o and Gemini 2.5, the framework enables sophisticated, lean reasoning without the need for massive custom-labeled datasets. Pilot validation achieved a 0.946 mAP in distinguishing state-dependent equipment loads. By formalizing the visual signatures of waste, the LCVT establishes the data infrastructure necessary for proactive, VLM-driven decision support in construction AI.