3D Pipeline Intelligent Detection Framework

• The paper introduces a computational framework that integrates multi-view feature extraction and physics-informed validation with deep learning modules to accurately detect and localize pipeline structures.
• It employs cross-dimensional techniques like DCO-YOLO and CEPDNet, achieving up to 96.2% accuracy and sub-millimeter defect localization through effective data fusion and spatial association.
• Experimental results in urban GPR and underwater imaging demonstrate enhanced performance, real-time inference speeds, and significant improvements over conventional detection methods.

A 3D pipeline intelligent detection framework is a computational system designed for accurate, robust, and efficient recognition and spatial localization of pipeline structures in volumetric environments. This paradigm encompasses multi-view feature extraction, cross-dimensional deep learning enhancements, geometry-aware association algorithms, and fusion strategies tailored to the physical sensing modalities, such as ground-penetrating radar (GPR) images or structured-light underwater imaging. The goal is to resolve ambiguities and improve small-scale object recognition via multi-source data integration and spatial reasoning. The following sections delineate the key methodological components, algorithmic innovations, and demonstrated empirical advantages, with grounding in recent research (Lv et al., 24 Dec 2025, Hu et al., 12 Dec 2025).

1. Multi-View Feature Analysis and Physics-Informed Model Validation

Pipeline detection in complex media, such as urban subsurfaces or underwater environments, is intrinsically challenged by weak inter-view correlations, clutter, and low SNR for small-diameter targets. The framework leverages three canonical GPR views—B-scan (x–t), C-scan (x–y), and D-scan (y–t)—to extract complementary feature curves interpretable as $$\mathbf{f}(x,y,t)=[f_B(x,t),f_C(x,y),f_D(y,t)]^\mathsf{T}\in\mathbb{R}^3$$ (Lv et al., 24 Dec 2025). These features are cross-validated between forward FDTD simulation $\mathbf{f}_{\rm sim}$ and field measurements $\mathbf{f}_{\rm meas}$, enforcing a high normalized inner-product correlation $$\rho=\frac{\langle \mathbf{f}_{\rm sim},\mathbf{f}_{\rm meas}\rangle}{\|\mathbf{f}_{\rm sim}\|_2\|\mathbf{f}_{\rm meas}\|_2}>\tau_{\rm corr}$$, with $\tau_{\rm corr}\approx0.9$ for signature alignment.

In underwater pipeline inspection, multi-mode structured light imaging employs translation, rotation, and combined motion strategies to scan pipelines with adaptable spatial and angular resolution (Hu et al., 12 Dec 2025). A rapid distortion correction (FDC) model efficiently rectifies underwater images, and extrinsic calibration is solved via factor graph optimization aligning optical and acoustic sensor frames with sub-decimeter error.

2. Deep Learning-Enhanced Detection: DCO-YOLO and CEPDNet Modules

The detection backbones are extended from canonical YOLOv11 by integrating three cross-dimensional modules:

• DySample: A dynamic, learnable upsampling grid $X'=\mathrm{grid\_sample}(X,S)$ with offset field $O$ improves the recovery of fine edge structures in GPR signatures.
• CGLU: Convolutional Gated Linear Units spatially modulate activations via pointwise gate $A=\mathrm{GELU}(\mathrm{DWConv}(W_g X))$, focusing model capacity on pipeline edges for enhanced discrimination.
• OutlookAttention: Local window self-attention stengthens the network’s ability to relate central hyperbola vertices to peripheral diffraction features via $$Y_{i,j}=\sum_{(u,v)\in\mathcal{N}_{i,j}} \widetilde{A}_{i,j}(u,v)V_{u,v}$$.

For underwater scenes, context-enhanced pipeline edge detection networks (CEPDNet) utilize residual, dilated, and self-attention blocks to output keypoint heatmaps distinguishing pipeline contours from the background. This enables robust input filtering for subsequent 3D reconstruction stages, achieving 99.04% accuracy and 219 FPS inference speed (Hu et al., 12 Dec 2025).

3. Spatial Association via 3D-DIoU Matching and ED-ICP Registration

Ambiguity reduction across views is achieved by lifting orthogonal 2D detections into 3D cuboids and matching them with a geometric metric, the 3D-DIoU (Lv et al., 24 Dec 2025):

$$\mathrm{3D\text{-}DIoU}(A,B) = \mathrm{3D\text{-}IoU}(A,B) - \frac{d^2}{\ell^2}$$

where $d$ is the centroid distance and $\ell$ the diagonal of the minimal enclosing box. Only triples with $\mathrm{3D\text{-}DIoU}>\tau$ for all pairs are fused into valid pipeline hypotheses, unifying B/C/D scan results into coherent 3D instances.

In underwater imaging, edge detection-based ICP (ED-ICP) aligns partial stripes or sub-clouds iteratively, weighting local correspondences by edge confidence and normal consistency. The cost is

$$C(R,t)=\sum_{i=1}^{N} w_i \| R p_i + t - q_{j(i)} \|^2$$

with convergence via SVD and adaptive thresholds. AEKF-fused pose estimates provide initial alignment, and the combination yields framewise RMSE as low as 1.08 mm and global merging error of 1.03 mm over hundreds of millimeters of pipeline length (Hu et al., 12 Dec 2025).

4. Data Fusion Strategies and Multi-Mode Operation

The pipeline supports data-efficient mid-level fusion:

• Multi-frequency pose fusion: Hierarchically combine high-rate IMU, mid-rate actuator, and low-rate camera/DVL data to stabilize spatial localization.
• Perception fusion: Ensures consistency between stripe-plane estimation (structured light) and pose estimates for accurate 3D back-projection in underwater mode.

A multi-mode imaging strategy allows translation-only, rotation-only, or combined scanning, addressing geometric variability and providing sub-millimeter defect localization regardless of pipeline orientation or occlusion.

5. Empirical Results, Ablation, and Comparative Insights

In complex multi-pipeline urban GPR scenarios, the DCO-YOLO pipeline achieves accuracy 96.2%, recall 93.3%, and mAP 96.7%, outperforming a YOLOv11 baseline by up to +2.1 pp in recall. Ablation reveals additive gains from DySample (+0.6 pp recall), CGLU (+0.5 pp), and OutlookAttention (+0.4 pp), with full module synergy validated by Grad-CAM++ heatmaps sharply localized on pipeline edges.

For underwater structured-light systems, the UW-SLD 3D pipeline demonstrates maximal imaging errors below 3% across operational modes, extrusion calibration errors below 4 cm, and defect localization error less than 1.22 mm over 0.5 m pipeline segments. Against Zhang (2000) distortion correction or prior light-plane methods, reprojection errors and X-Y-Z spatial error are reduced by 6–80% (Hu et al., 12 Dec 2025).

Table: Core Quantitative Outcomes

Domain	Method	Accuracy (%)	Recall (%)	mAP (%)	Length Error (mm)
Urban GPR	DCO-YOLO	96.2	93.3	96.7	—
Underwater	UW-SLD + ED-ICP	—	—	—	1.22 (0.24%)

6. Limitations and Future Directions

Current frameworks assume static scenes—in dynamic contexts (e.g., vibrating or mobile pipelines), depth-based registration and mask segmentation yield artifacts. For GPR, ambiguities persist when multiple pipelines are tightly clustered or exhibit weak signatures. The system requires high-bandwidth links (cloud backend for SLAM), and environmental constraints limit performance (indoor lighting for RGB-D, water turbidity for structured light).

Planned extensions include:

• Learned SLAM backends for drift reduction;
• LiDAR+RGB fusion for outdoor scenarios;
• On-device lightweight fusion for edge deployments;
• Fine-tuning of edge detectors for highly turbid underwater or multi-frequency radar conditions.

7. Impact and Application Spectrum

Demonstrated applications range from urban underground pipeline mapping, leakage inspection, AR/VR asset creation, robotic scene graphs, to autonomous underwater defect quantification. The integration of physical model validation, cross-dimensional deep learning, and geometric reasoning delivers substantial gains in accuracy, reliability, and deployment efficiency for industrial and engineering tasks—enabling scalable, real-time 3D pipeline intelligence in challenging operational environments (Lv et al., 24 Dec 2025, Hu et al., 12 Dec 2025).