Accept: 2026-07-01 / Publish: 2026-07-14

Accurate localization of occupants in complex indoor facilities and underground spaces is essential for effective emergency response. However, conventional closed-circuit television (CCTV) and infrared (IR) sensor-based systems suffer from severely degraded detection performance under smoke, reduced illumination, and thermal noise, while also raising privacy concerns. To overcome these limitations, this paper proposes an indoor human detection and geo-referencing system employing distributed 60 GHz millimeter-wave (mmWave) radar sensors that are robust to adverse environmental conditions and inherently privacy-preserving. The proposed system deploys 14 distributed sensors—combining unidirectional (1-Way) and omnidirectional (4-Way) configurations tailored to spatial characteristics—and processes point cloud data in real time through a lightweight middleware architecture. Raw measurements are transformed from spherical to Cartesian coordinates, after which a dimensionality-reduction-based algorithm separates height information, performs two-dimensional density-based clustering (DBSCAN), and subsequently restores height to maximize computational efficiency. A geo-referencing technique incorporating sensor installation positions and orientations precisely converts sensor-relative coordinates to a world coordinate system aligned with building floor plans. The proposed system was validated through a 60-day field demonstration at the Daejeon World Cup Stadium, achieving 99.96% uptime. A controlled geo-referencing evaluation using a publicly available paired radar–skeleton dataset—which provides synchronized mmWave radar point clouds and optical motion capture-based 3D ground-truth coordinates—yielded a mean absolute error (MAE) of 0.59 m against the external skeleton reference after virtual geo-referencing alignment, while an image-based baseline employing a fixed-height assumption exhibited an MAE of 6.04 m, representing an approximately 10.2× improvement. These results demonstrate the practical viability of the proposed framework for privacy-preserving indoor occupant monitoring in complex indoor environments.

학회기간: 2026-07-01 ~ 2026-07-03

본 논문에서는 SHP 기반 실내공간을 정량적으로 분할하고, 격자 중심 연결 방식과 자유 좌표 연결 방식을 이용하여 실내 이동 네트워크를 구축·편집하는 시스템을 제안한다. 제안 시스템은 PyQt5, GeoPandas 및 Shapely를 기반으로 회전 보정 공간분할, 네트워크 생성·분할, 속성 편집 및 SHP 저장 기능을 제공한다. 공공기관 등 13개 실내공간 데이터에 적용하여 동일한 환경에서 공간분할과 네트워크 작성이 가능함을 확인하였다. 제안 시스템은 SHP 기반 실내 네트워크 구축을 지원하고, IndoorGML 구축을 위한 기초 공간자료 작성 도구로 활용할 수 있다.

Accept: 2026-06-10 / Publish: 2026-06-22

Geographic knowledge graphs (GeoKGs) have attracted growing attention for urban digital-twin systems, yet prior work has primarily targeted 3D-visualization efficiency on small scenes rather than district-scale infrastructure analysis. This paper proposes a declarative GeoKG framework that redirects the focus toward analytical capabilities at district scale, with three contributions. First, a declarative rule engine of 7 matching strategies and 18 rules automatically generates 936,739 relationships from 118,231 nodes on Yuseong-gu (Daejeon) and 1,402,275 from 320,863 nodes on Sejong, with 88–100 % independent cross-validation precision for attribute- and proximity-based rules. Second, a physical road-network topology built from TN_RODWAY_NODE/LINK and TL_SPRD_MANAGE delivers 100 % topological coverage of the building–road linkage on both cities, with entrance-aware FRONTS_ROAD anchoring 98.4 % of buildings and the underlying positional quality decomposed quantitatively in the precision-validation section. Third, 16 graph-based analysis functions—including safety assessment, dead-zone identification, and road-closure impact simulation—support evidence-based urban management. End-to-end Neo4j builds complete in ~6 min on Yuseong and ~22 min on Sejong, with sub-2 % rule-engine run-to-run variance across ten independent builds per region; the Sejong/Yuseong engine ratio (3.11  × ) closely tracks the 2.92  × ratio of underlying pair operations, confirming near-linear scaling. Because the rules are declarative JSON specifications independent of any region, the framework is portable to other cities providing equivalent open data. Case studies reveal infrastructure inequality across 45 stable legal-dong of Yuseong (safety-score range 18.5–79.8, 4.3  × best-to-worst gap) and 131 stable legal-dong of Sejong (30.4–80.0, 2.6  × gap), with the gap concentrated in suburban areas with sparse object-address shelter coverage.

Accept: 2026-05-06 / Publish: 2026-05-21

Geographic question answering (QA) requires factually accurate responses grounded in structured spatial knowledge. However, existing approaches based on large language models (LLMs) suffer from factual hallucination, while retrieval-augmented generation (RAG) lacks structural reasoning capabilities. We propose GNLM (Graph-Native Language Model), a novel architecture that restructures the traditional LLM pipeline by performing all reasoning through knowledge graph traversal and restricting the role of LLMs to natural language generation. The knowledge graph is constructed exclusively from government public data sources (building registry, road name addresses) and expert-curated domain knowledge, ensuring factual reliability without LLM-generated content. GNLM introduces four key contributions: 1) a Graph-First, LLM-Last architecture with fact validation that structurally prevents fabrication (all observed errors stem from knowledge gaps, not false generation); 2) an extended Resource Description Framework (RDF) triple attribute model that represents entity properties as independent graph nodes connected by semantic edges, enabling property-level querying and meta-reasoning; 3) coordinate-free spatial reasoning that leverages the Korean road numbering standard (20 m interval, k=10  m per unit) for proximity ranking without Global Positioning System (GPS) coordinates; and 4) a 14-strategy reasoning router that classifies queries into specialized graph traversal algorithms. Evaluated on a real-world dataset of Daejeon, South Korea ( 58,882 nodes, 80,088 edges from 58,408 buildings and 711 roads), GNLM achieves 88.4% accuracy on 250 diverse queries with a hallucination rate (HR) of only 5.7% on 35 fact-verification queries, compared to 60.0% for pure LLM, 25.7% for Dense RAG, and 17.1% for GraphRAG—demonstrating a consistent improvement as structural grounding increases. Cross-model validation with llama3.1:8b confirms LLM-agnostic generalizability. Multi-region validation on Sejong City ( 17,008 buildings) with identical methodology confirms cross-region generalizability (Yuseong-gu avg ρ=0.976 , Sejong avg ρ=0.973 , both over 10 roads). Road-number-based proximity ranking achieves Spearman ρ=0.9878 and Precision@ 10=96.0% against GPS ground truth, with short-range (<500 m) Mean Absolute Error (MAE) of only 31.9 m, demonstrating that structured address numbers alone can effectively preserve spatial ordering without any coordinate data. Cross-road queries, supported via a graph-native intersection topology with 34 annotated intersections, achieve a 90% resolution rate with Spearman ρ=0.701 —substantially lower than same-road accuracy, indicating a current practical limitation that motivates the multi-hop and automated intersection-detection extensions.

Accept: 2026-04-22 / Publish: 2026-05-06

Large language models (LLMs) have demonstrated considerable capability in spatial reasoning, yet prior work has focused primarily on diagnosing the limits of LLM spatial cognition rather than identifying which input representations improve performance. This study addresses a complementary question: how should spatial information be structured to support LLM-based spatial reasoning? We propose a five-level spatial abstraction hierarchy—from raw coordinates (L1) through semantic labels (L2), natural-language relations (L3a), domain-specific rules (L3b), to pre-computed distance tables (L4)—and evaluate its effects on four task families: category-filtered retrieval (H1), relative position reasoning (H2), rule-constrained inference (H3), and distance comparison (H4). Experiments span controlled synthetic urban scenes (8, 12, and 16 objects), four real-world geographic information system (GIS) scenes derived from OpenStreetMap (Daejeon, South Korea), four open-source models, and two commercial models (GPT-5.2, Claude Sonnet 4.5), with paired bootstrap confidence intervals, McNemar tests, and five-seed replication. The results comprise one design precondition validation and three principal findings. As a precondition, semantic labels are confirmed as a structural prerequisite for category-based tasks—an expected but empirically verified requirement ( p<0.001 ). The three findings are: 1) explicit natural-language relation sentences interfere with coordinate-based position reasoning ( −19.7±7.0%p ), with transition analysis confirming that 83% of answer flips are correct-to-incorrect and conflict-trust experiments revealing a wrong-follow rate of 25.6%, although this interference is model- and context-dependent; 2) domain-specific rule injection provides conditional benefit modulated by scene complexity ( +18.3±4.9%p , with substantial scene-level variation); and 3) distance tables yield gains inversely related to the model’s numerical computation ability (+ 2.7 to +84.1%p ). These results establish that effective AI-ready spatial representations should be optimized for task relevance and model fit—prioritizing semantic grounding and task-contingent augmentation over information density.