Adaptive multi-sensor point cloud fusion for geometric reconstruction toward digital twin construction based on quality-driven weight optimization.

Digital twin technology demands high-fidelity three-dimensional geometric representations, yet single-sensor acquisition systems suffer from limited coverage, uneven sampling density, and environmental susceptibility. This paper presents an adaptive weight allocation framework for multi-sensor point cloud fusion that dynamically evaluates sensor reliability based on local geometric features and statistical consistency measures. Unlike conventional fixed-weight or rule-based approaches, the proposed methodology computes spatially-varying fusion weights through multi-dimensional quality assessment encompassing geometric consistency, point density, outlier characteristics, and surface normal coherence. An octree-based spatial partitioning strategy enables efficient processing while maintaining sufficient local samples for statistically meaningful quality estimation. Experimental validation across manufacturing, architectural heritage, and infrastructure scenarios demonstrates 15–46% improvement over baseline methods—28–42% in absolute positional accuracy measured against laser tracker references in the manufacturing scenario, and 15–46% in cross-validated internal consistency across all three scenarios—achieving mean positional errors below 2 mm and 96.7% coverage of sensor-accessible surfaces. Real-world deployment in manufacturing equipment monitoring and construction progress tracking detected dimensional deviations—specifically a 0.8 mm spindle drift and 12–27 mm pipeline installation offsets—that periodic manual caliper measurements and visual inspections had not flagged, validating practical utility for industrial 3D modeling applications. The framework’s sensor-agnostic design and computational efficiency make it suitable for near-real-time geometric model synchronization within digital twin pipelines across diverse operational environments. We note that the current weighting coefficients are empirically tuned rather than dynamically learned; nevertheless, sensitivity analysis confirms robustness to coefficient perturbation across the tested scenarios.