Multi-MLP Neural-Implicit SLAM for Real-Time UAV Inspection in Dynamic Aviation Environments

Real-time 3D mapping and localization for Unmanned Aerial Vehicles (UAVs) performing in-flight inspections of civil aviation infrastructure is crucial for ensuring safety, reducing downtime, and enabling predictive maintenance. However, existing SLAM frameworks such as ORB-SLAM3 and NICE-SLAM struggle with the dynamic conditions encountered during UAV flight, including rapid viewpoint changes, moving obstacles, and non-rigid scene elements. We propose DMN-SLAM-UAV, a lightweight neural-implicit SLAM system tailored for UAV inspection tasks that integrates a real-time dynamic-object segmentation front end, a hierarchy of five compact MLP decoders for multi-scale geometry and appearance residuals, and an incremental octree (i-Octree) for efficient map storage. The i-Octree enables sparse, log-time updates, allowing onboard execution on an NVIDIA Jetson Orin with an average per-frame latency of 48 ms. We evaluate DMN-SLAM-UAV on a newly curated Dataset of Aerial Inspection Sequences (DAIS)—20 flight trajectories over runways and fuselage sections—and benchmark against ORB-SLAM3 and NICE-SLAM. Our approach reduces absolute trajectory error (ATE) by 37.6% and improves reconstruction completeness by 23.4% compared to ORB-SLAM3, while achieving 42% faster mapping updates than NICE-SLAM. Ablation studies confirm the contributions of dynamic masking and the multi-MLP architecture with statistical significance (p < 0.01). Furthermore, we release the DAIS dataset and our open-source implementation to foster future research. This work paves the way for robust, scalable UAV-based inspection solutions in civil aviation and beyond.