AI damage detection lead
Overall project management, comparison of 4-Tier detection algorithms and selection of optimal models, multi-sensor fusion strategy, lightweight model implementation
SPECIAL & SPATIAL INTELLIGENCE
National R&D · National Disaster Management Research Institute · 2026 – 2028
Suspension footbridges, ziplines, bungee jump towers. We are developing a safety inspection platform in which commercial drones and AI reach adventure leisure facilities that people cannot easily approach, examine them, detect damage, and leave a record behind. In this project GIPTER is responsible for the drone platform and SDK-based autonomous flight control.
FIELD · Yeoju Challenge Park
Project overview
The Ministry of the Interior and Safety conducts joint government safety inspections of adventure leisure facilities, but these rely on expert visual inspection. The National Disaster Management Research Institute commissioned an academic research project to advance this into an intelligent inspection framework based on drones, AI, and big data, and an industry–academic consortium is carrying it out over three years from 2026.
Consortium roles
Overall project management, comparison of 4-Tier detection algorithms and selection of optimal models, multi-sensor fusion strategy, lightweight model implementation
Labeling rules, data QC and augmentation, drone–AI–DB integration architecture, risk grading and 3D Damage Map
Two-stage autonomous flight design, onboard control module, thin-line obstacle avoidance, triple-redundant hybrid navigation, sensor calibration, intelligent prototype
Why now
Members that sit high up and in tight places — cables, pylons, anchorages, connection joints — can only be covered so widely and so often when a person has to check them by eye. Changes that appear small at first, such as fine cracks, corrosion, fastening faults, and cable deflection, can be missed depending on image quality and the inspector's experience.
Adventure leisure facilities such as suspension footbridges, ziplines, and bungee jumps have grown rapidly nationwide, and the risk of safety accidents from aging has grown with them.
The institutional basis and lead agency needed to oversee installation standards and safety management practices are inadequate, so blind spots in safety management remain.
Areas that are hard to approach — upper pylons, cable mid-spans, hanger connections, anchorages — are difficult to inspect precisely at all.
Location-based records and time-series comparison of inspection results are difficult, which makes it hard to judge quantitatively how far damage has progressed.
What makes it different
Conventional drone inspection plans its routes on the assumption that drawings or a 3D model are already available. This platform removes that assumption. The operator specifies a location, and the drone recognizes the facility on its own, makes its decisions onboard, and keeps flying even where GNSS drops out.
No CAD or BIM is required. In the stage-one survey flight, LiDAR generates a 3D shape in real time, the main inspection areas are identified automatically, and precision data is then acquired. Immediate on-site operation is possible without a separate shape database.
The core decisions required for flight safety — obstacle identification, facility recognition, GNSS-denied response — are made in real time on the onboard computer (Manifold 3, up to 100 TOPS). Autonomous flight continues even during communication delays or network failures.
A three-layer navigation stack running from RTK/GNSS → LiDAR-inertial navigation (FAST-LIO2) → UWB absolute position correction. Position is held even under bridge piers, across valley spans, and in areas dense with metal structures; if every layer fails, the drone hovers immediately and returns automatically.
RGB alone struggles to tell shadows, backlight, moisture stains, and similar paint colors apart from damage. EO, IR, and LiDAR divide the work and cross-validate one another, reducing false detections systematically.
Damage and non-damage data acquired during inspection operations accumulates automatically into the training dataset. Anomaly detection trained on normal data (reverse learning) works around the initial shortage of data, creating a virtuous cycle that improves with use.
Where full autonomy is difficult, the inspector's judgment is built into the procedure. For hazardous vertical structures such as bungee jump towers, operator approval is required at every segment transition, and clicking a POI on screen sends the drone into an automatic orbit for imaging.
Autonomous flight · GIPTER's part
A flight architecture built on the DJI Matrice 400 RTK, combining MSDK v5 (ground station) and PSDK v3 (onboard). The drone first circles the facility's exterior to acquire its shape, then automatically generates and executes member-level precision inspection routes from that shape. Stage two is divided again into two layers, Macro-Path and Micro-Path, where the onboard AI judges each image immediately after capture and fine-tunes its position.
OUT → 3D point cloud · safety zones · priority inspection areas
Preprocessing→segmentation→safety zones→
waypoints→TSP optimization→KMZ
GNSS ▸ LIO ▸ UWB three-layer navigation + safety_monitor_node watches the entire route · on an anomaly, SEMI_AUTO_HOLD → FAILSAFE
Flight strategy by facility type
| Facility | Flight mode | Automation level | Core algorithms |
|---|---|---|---|
| Suspension footbridge | Linear + Facade | Fully automatic | Grid (lawnmower) scan of the bridge deck · main cable catenary (y = a·cosh(x/a)) following · zigzag facade scan of the pylons (overlap 80% longitudinal / 60% transverse) · automatic POI addition at major connection joints. The whole route is generated from span length, bridge width, and cable height difference alone |
| Zipline | Cable tracking, 4 stages | Fully automatic | ① A preliminary flight detects the cable automatically → ② 3D centerline extraction from the point cloud plus catenary fitting (RANSAC + Levenberg-Marquardt) → ③ precision following at a fixed clearance, switching to an orbit on reaching a support → ④ automatic analysis of deflection against the design reference line, flagging segments beyond tolerance and re-flying them with NBV |
| Bungee jump tower | POI + vertical scan | Semi-automatic · human-in-the-loop | The tower is divided into equal segments for a level-by-level facade scan, with operator approval at every segment transition · bungee cord anchors and platform joints are designated as POIs by clicking on screen → automatic orbit (radius, altitude, and speed adjusted in real time, gimbal tracking on a 50 ms cycle) |
6-step path generation · path_generator_node (ROS 2 Action)
Voxel grid 0.05 m downsampling → statistical outlier removal → RANSAC ground separation → ROI filter. Data volume compressed by about 80%, 32 → 6 ms/frame
Euclidean clustering → PCA linearity and planarity → classification into cable / deck / pylon → point-to-line cable-specific extraction
A SafetyZone per member. EO clearance 3.0 m, LiDAR 5.0 m by default. For thin-line members, actual radius + angular resolution × distance + wind-load margin
Positions computed from surface normal × clearance distance, gimbal yaw/pitch calculated automatically, additional points inserted where FOV overlap falls below 80%
TSP visit-order optimization (Google OR-Tools CP-SAT) → battery consumption prediction → automatic insertion of mission split points
WPML (waylines.wpml + wpmz/res) → KMZ package. Loads directly in DJI Pilot 2 and in our own GCS app
AI damage detection · 4-Tier pipeline
Damage on adventure leisure facilities varies widely in size and form, and ground-truth data is often insufficient. The system first locates the member, then analyzes its surroundings in detail by damage type, quantifies deflection and deformation with LiDAR, and catches unlabeled damage as an anomaly relative to the normal state. Fast first-pass screening (onboard) and precise interpretation (offboard) are kept separate.
Quickly locates pylons, cables, hangers, decks, and connection joints in drone imagery. Speed-first and precision-first models are compared by mAP and by inference time on Manifold 3.
Cracks are handled by pixel-level segmentation (mIoU, Dice), early-stage corrosion and paint delamination by reconstruction-based anomaly detection trained on normal images alone, and missing bolts by a detector plus a classifier.
Cable deflection, pylon verticality deviation, and member curvature are computed from LiDAR and image geometry. Viewpoint distortion in 2D imagery is corrected against the LiDAR 3D reconstruction as the reference frame.
Detects damage with scarce labels and unforeseen anomalies in a zero-shot or few-shot manner. Methods that determine damage from a text prompt alone are also compared.
Multi-sensor false-positive suppression
A DJI Zenmuse H30T (48 MP wide + 40 MP zoom, 10× optical and 200× hybrid, 1280×1024 thermal) and a Zenmuse L2 (905 nm, ≥ 100 pt/m² at 100 m altitude) are carried on one airframe, and the sensors are combined according to the damage type. For very small damage such as hairline cracks, hybrid 20× zoom secures an equivalent GSD from 4 m instead of approaching to 1 m with the wide lens.
| Cause of false detection | Sensor · strategy in response |
|---|---|
| Shadow · backlight | RGB supplemented by multiple capture angles and adjusted time of day, with IR thermal imaging providing auxiliary temperature-based anomaly detection independent of illumination |
| Moisture · oil stains | IR temperature-difference patterns distinguish oil leaks and moisture traces from actual damage, with cross-validation |
| Similar paint colors | LiDAR surface roughness information combined with multi-angle RGB capture to raise the recognition rate |
| Geometric misjudgment from viewpoint distortion | The LiDAR 3D reconstruction is used as the reference frame to correct shape judgment errors in 2D imagery |
| Connection joint · rope degradation | IR provides auxiliary detection of thermal anomalies such as abnormal heating in the trolley system and internal degradation of elastic ropes |
Training data
Damage data specific to adventure leisure facilities is almost absent from public datasets. Three paths — on-site measurement, public data, and simulation synthesis — run in parallel across all project years, and quality is kept consistent through standard labeling rules and automated QC.
Drone imagery captured at actual facilities. It is collected in parallel throughout all project years, and during flight the onboard system classifies images in real time into four quality grades, automatically sending a re-capture alert for any segment whose Grade D share exceeds 30%.
Existing datasets such as AI Hub bridge safety inspection drone imagery (about 4,200 crack images), Kaggle Concrete Crack, and KITTI are preprocessed and reused to accelerate initial training in year 1.
Damage textures are composited onto 3D models, and illumination, distance, angle, and blur are varied to generate diverse damage patterns. In year 3, adverse-condition data such as low light, strong wind, and low resolution will be built up intensively.
Actual facilities without drawings were reconstructed through LLM-based 3D modeling. The zipline tower at Yeoju Challenge Park was modeled from 125 on-site photos and videos, and the bungee jump tower from publicly available drone footage alone. A Blender + ArduPilot virtual drone is then flown in these models to render drone-viewpoint images while varying flight path, focal length, and resolution, and real HDRI backgrounds are applied to produce synthetic training data.
A hierarchical label scheme that maps one-to-one onto the Tier structure of the AI pipeline. RGB is recorded as bounding box / polygon / polyline, IR as a pixel-level thermal mask, and point clouds as 3D box / point-wise segmentation. The scheme is being formalized as an open public guideline in preparation for distribution through AI Hub.
Bounding box IoU ≥ 0.85, polygon IoU ≥ 0.80, 3D box IoU ≥ 0.70. Agreement of ≥ 90% between two or more annotators on the same image (Cohen's κ ≥ 0.80). We run a three-step procedure — initial labeling → cross-review → approval by a facility expert — and maintain a Hard Case Dictionary.
Blur judgment based on Laplacian variance (threshold 100 — in a preliminary experiment on 500 actual drone images, 97.8% of visually blurred images were classified correctly), Brenner gradient as a secondary metric, and checks for exposure, resolution, and SSIM duplication. Dataset versions, collection paths, and QC history are tracked per batch as metadata, and HTML reports are generated automatically.
Data flywheel
Images acquired on inspection flights pass through onboard quality grading and accumulate automatically as training candidates. Detection results confirmed by the inspector become labels, and data that passes QC is fed into retraining. Operational history becomes model performance.
Platform stack
A feasibility-first principle: first confirm the SDK support range of the commercial drone the institute operates, then choose the design that draws the most performance from within it. Instead of the discontinued OSDK, the extended flight control API of PSDK v3 is used to control flight and sensors together onboard.
3-year roadmap
The three axes of drone flight control, AI detection, and training data development run in parallel from year 1, in a staged integration structure where each year's deliverables become the basis for the next.
3rd-year field validation · KPI design
Measurement of elapsed time, error counts, and data loss rate at each stage, from flight → AI detection → inspector confirmation → database storage → report generation
Measurement of detection performance degradation, navigation switchover success rate, and obstacle avoidance rate in low light, strong wind, and GNSS-denied segments
Comparison of inspection time, number of defects detected, and detection rate for damage in inaccessible areas at the same facility
SUS surveys and in-depth interviews with inspectors and managers to identify pain points and improvement priorities
Expected impact
The goal of this project is to build a preemptive, preventive inspection framework for adventure leisure facilities that sit in a safety management blind spot, and to extend the validated operating model to public facilities of local governments nationwide.
Drone-based automatic inspection replaces high-elevation and close-approach work that risks human injury, and shifts management onto quantitative data.
Three sensors (EO, IR, LiDAR), three navigation layers (GNSS, LIO, UWB), and fail-safe. Precision and safety are secured together.
The hierarchical standard labeling rules for facilities, members, and damage, together with the datasets, will be released publicly to provide a basis for follow-up research and industrial adoption.
The same architecture extends to the inspection of thin-line and high-elevation structures such as bridges, transmission towers, wind turbines, and tower cranes.
Let's talk
This page is based on the proposal, work plan, design deliverables, and interim report of the National Disaster Management Research Institute academic research project “Development of a commercial-drone-tailored facility safety inspection platform and management technology”. The figures and parameters are year-1 design reference values and will be confirmed after field validation.
We use Google Analytics cookies to understand how our website is used. With your consent, they help us see which pages are read, as anonymous statistics. They are not used for advertising. Privacy Policy