Slide

3D Recon V2

Detailed subsea models for inspection and measurement Assessing subsea assets requires detailed information about their condition and geometry.

3D Recon V2

Detailed subsea models for inspection and measurement Assessing subsea assets requires detailed information about their condition and geometry.

3D Recon V2

Detailed subsea models for inspection and measurement

Assessing subsea assets requires detailed information about their condition and geometry. Zupt’s 3D Recon V2 produces high-resolution, geospatially accurate 3D models for inspection, measurement, and integrity management.

3D Recon V2 is a subsea stereo imaging system with two high-resolution machine vision cameras and a high-performance MEMS IMU. The system is housed in a compact subsea enclosure rated to 4,000 meters.

By combining inertial technology with imaging sensors, it generates real-time point clouds for navigation and quality control, alongside high-density models for detailed inspection and engineering assessment.

Applications

Originally developed for Integrity Management (IM) and Inspection, Repair, and Maintenance (IRM), 3D Recon V2 is also used for:

Jumper and spool metrology
Mooring chain-link inspection
Hull inspection
Pipeline out-of-straightness (OSS) surveys
Pre-installation and as-built documentation
Real-time positioning relative to structures
Real-time object detection and pose estimation of underwater objects

How 3D Recon V2 Works

3D Recon V2 combines 3D reconstruction, photogrammetry, computer vision, and inertial navigation to calculate camera positions and orientations and generate high-resolution point clouds.

The system supports real-time processing and optimized offline processing.

Real-Time Point Cloud Generation

The real-time process provides navigation, trajectory estimation, and point cloud generation while data is being collected.

Alignment and navigation: IMU data determines sensor orientation, including heading, pitch, and roll. The navigation engine integrates acceleration and angular velocity to update position, velocity, and attitude.
Feature detection: Cameras capture images at 5–10 Hz. Features are identified using intensity variations.
Feature matching: Descriptor vectors are calculated and used to match features between the stereo cameras and sequential images. This creates a sparse point cloud with approximately 5–10 cm pixel resolution.
Map integration: Features are matched against the existing map and projected from the camera frame into the global frame. RANSAC and optimization checks assess consistency with the map.
Navigation updates and path optimization: Successful matches update the navigation state and covariance. New features are added to the map, and keyframes are collected for map optimization when specific events occur.
Dense point cloud generation: Calculated positions and orientations are used with disparity maps derived from light-intensity comparisons between the left and right cameras to generate high-density point clouds.
SLAM and loop closure: Real-time processing allows the operator to see their position relative to the structure, track their path, and assess coverage. Revisiting previously mapped areas triggers loop closure, correcting accumulated drift in camera pose and trajectory to produce globally consistent dense point clouds.

Offline Dense Point Cloud Generation

Offline processing uses higher-quality image features and additional matching and optimization steps to generate detailed models.

Feature detection: Features are extracted from every captured image, using higher-quality features than the real-time process.
Feature matching: Descriptor vectors are used to match features across the collected images, including non-sequential images.
Structure from Motion: Structure from Motion and Bundle Adjustment estimate feature positions and camera poses for each image, producing a sparse 3D point cloud and the corresponding camera poses.
Dense point cloud generation: Calculated camera poses and disparity maps are used to generate high-density point clouds. Disparity maps are created by comparing light intensity between patches around sparse features across the images.
Mesh generation: Surface normals are calculated using each point’s three nearest neighbors to define a local plane, as part of generating a detailed 3D model of the structure.

Real-Time Object Detection and Pose Estimation

3D Recon V2 also supports trained object detection and pose estimation.

Model training: Deep learning models are trained using collected and annotated data to recognize a specific object and generate bounding boxes and prediction confidence scores.
Keypoint prediction: The model and processing pipeline identify specific keypoints for pose estimation.
Pose estimation: Correspondences between the 2D keypoints and known object dimensions are used with the Perspective-n-Points algorithm to calculate the object’s pose relative to the cameras.

Inspection and Measurement Capabilities

Spatial scaling for measurements

The tightly coupled inertial navigation solution provides accurate spatial scaling within the delivered dense models, supporting linear, angular, and area measurements.

Change detection

Models generated by 3D Recon V2 can be compared with historical models to automate change detection. This allows operators to monitor structural changes over time and support integrity management decisions.

Additional sensor integration

Data from other sensors, including contactless cathodic protection sensors and hydrocarbon sniffers, can be integrated to produce a combined model or heat map for structural integrity analysis.

Deliverables

3D Recon V2 provides:

High-resolution, geospatially accurate 3D models with submillimeter pixel resolution
High-resolution still images
Conventional video
Real-time 3D models for verification
Offline 3D models for detailed inspection

These outputs support measurement-based assessments and evaluation of structural condition over time, and can be incorporated into existing workflows.

Data Acquisition and Integration Requirements

A survey mux is required to integrate the 3D Recon V2 spread. Power, serial communications, and image data connect to the mux/ROV through a single PBOF cable.

The system requires: