Semantic–geometric asset mapping for fire-safety compliance inspection with multi-label state diagnosis

Operational fire-safety compliance inspection requires persistent asset-level records beyond equipment detection in images. Existing automated approaches mainly address asset presence or isolated subtasks, offering limited support for the location- and condition-aware records required in practice. This paper formulates fire-safety compliance inspection as a structured semantic–geometric estimation problem. The proposed output representation, termed a Semantic–Geometric Asset Map , consists of persistent asset nodes that jointly encode location, identity, and a multi-label belief over compliance-relevant states. This formulation exposes a key challenge: compliance evidence is often concentrated in small, attribute-specific regions of an asset, creating a sub-asset semantic granularity mismatch between whole-object recognition and fine-grained condition diagnosis. To address this, we introduce RiskEye , a task-specific multi-label condition estimator that aligns attribute prototypes with spatial visual evidence, enabling each compliance attribute to be inferred from its relevant local cues. We instantiate the formulation as a proof-of-concept inspection pipeline for RGB-D patrol observations acquired with a handheld or body-mounted sensor, combining ORB-SLAM-based tracking, RiskEye-based diagnosis, and recursive log-odds evidence fusion. On a curated multi-label diagnostic benchmark, RiskEye achieves 88.0% mAP and the most favourable accuracy–efficiency trade-off at 25.9M parameters and 4.20 GMACs. In a controlled field study comprising 120 walk-by inspection routes, the frozen end-to-end pipeline achieves 70.4% Joint Record Correctness. These results support reframing operational fire-safety compliance inspection from a frame-level perception problem into a record-centric estimation problem, in which repeated patrol observations are progressively fused into persistent semantic–geometric evidence for asset-level decision making.