DI-MAPS

DI-MAPS is a PC application that automates the conversion of 2D vessel drawings into machine-readable structural priors for autonomous drone (UAS) inspection of confined maritime spaces such as ballast water tanks. The solution is built as a four-stage pipeline: an Ingestion Layer that uses libraries to strip non-structural elements from vessel drawings and normalize geometry into a unified ship-reference coordinate system in SI meters; an Orthographic Fusion Engine that reconstructs a 3D volumetric model from 2D plan and section/profile projections using a space-carving algorithm, marking voxels as occupied only where structural lines align across corresponding views; a Semantic Annotation Layer that applies a YOLO model trained on engineering symbols to detect points of interest, calculate their 3D center-points, and assign them as spatial anchors to counter drone sensor drift; and an Export Layer that serializes the resulting model into standardized, drone-ready format. The output is a reviewed structural prior that supports exploration, expected-structure comparison, semantic anchoring, and relocalization for downstream UAS navigation and mapping systems.

DI-MAPS is a PC-based software application that converts 2D vessel engineering drawings into structured, machine-readable map priors for autonomous drone (UAS) inspection of confined maritime structures such as ballast water tanks. It addresses a key bottleneck in aerial robotic mapping: these environments are GNSS-denied, low-light, confined, and visually repetitive, which degrades SLAM performance and increases the risk of localization drift and mission failure. Ship drawings already contain rich structural information such as frames, bulkheads, manholes, ladders, pipes, and access features. However, today this information is extracted manually by experts, a process that is slow, inconsistent, and hard to scale. DI-MAPS automates this extraction and packages the result for direct use by drone navigation and mapping software.

The solution is organized into four processing components. The Ingestion Layer accepts vessel drawings, uses libraries to strip non-structural content, and normalizes all geometry to a common Ship-Reference Origin, converting units to SI meters for 1:1 real-world scale.

The Orthographic Fusion Engine reconstructs a 3D volumetric model from these 2D projections using a space-carving algorithm: a voxel at (x, y, z) is marked “occupied” only when structural lines are present at the corresponding (x, y) location in the plan view and (x, z) or (y, z) location in the section/profile views. This bridges the gap between flat, multi-view drawings and a coherent 3D structural skeleton without requiring true 3D CAD models.

The Semantic Annotation Layer applies a YOLO object-detection model, trained on a dataset of engineering symbols, to locate points of interest within the 2D orthographic views. Detected bounding boxes are converted into 3D center-point coordinates and stored as uniquely labelled spatial anchors, which drones can use as visual reference points to correct sensor drift during flight.

The Export Layer serializes the fused, annotated model into standard, interoperable formats: a compressed 3D Octomap for volumetric navigation, 2D occupancy grids per deck level for planar planning, and a semantic JSON manifest listing landmark IDs, types, coordinates, and detection confidence. All outputs carry full data provenance, including confidence scores and expert-review status, and are designed for plug-and-play integration with existing UAS SLAM and inspection pipelines.

The full pipeline is wrapped in a user-facing PC application that lets a maritime domain expert review, correct, or manually add landmarks before export, preserving human oversight over automated detections while keeping the overall workflow repeatable and auditable. The target maturity is TRL 7, demonstrated through an end-to-end run on real or representative ballast-tank drawings, with the exported map package validated against at least one UAS mapping/inspection workflow, simulator, API, or schema checker.