LDF-SLAM: lightweight decoupled architecture-based robust localization for UAVs in indoor constrained environments
TL;DR
This work contributes a new feature processing paradigm and a fusion constraint design strategy for robust pose estimation under weak-texture degradation in Three Dimensional (3D) imaging and embedded vision applications.
Abstract
Most simultaneous localization and mapping (SLAM) systems can’t work well in small places like warehouses, no GPS there, it’s too plain for texture, and not much field of view makes tracking worse. We present Lightweight Decoupled Fusion SLAM (LDF-SLAM) for Unmanned Aerial Vehicle (UAV) localization under such situations. The core is Feature-Enhanced Decoupled Semantic Keyframe Selection (FED-SKS), which separates the geometric tracking from semantic processing through triggering semantic segmentation only on motion-selected keyframes, and then using optical flow to propagate dynamic masks to non-keyframes. LDF-SLAM also has a multi-source fusion interface that allows extending with shelf geometry constraints as well as Pedestrian Dead Reckoning/Zero Velocity Update (PDR/ZUPT) factors. On Technische Universität München (TUM) RGB-D dataset experiment shows 100% tracking success rate and lower Root Mean Square Error (RMSE) than ORB-SLAM3 on translation-dominated sequence. This work contributes a new feature processing paradigm and a fusion constraint design strategy for robust pose estimation under weak-texture degradation in Three Dimensional (3D) imaging and embedded vision applications.