Scene recognition-based adaptive map switching for resource-constrained robotic navigation

초록

Balancing environmental complexity and computational efficiency remains a challenge in autonomous robot navigation. Conventional approaches use either two-dimensional (2D) or three-dimensional (3D) mapping. Although 2D projections fail to represent elevation and overhanging structures, 3D maps impose computational and memory burdens. To address this trade-off, an adaptive mapping framework, termed Adaptive Map Switching (ADAMS), was introduced. ADAMS enables real-time transitions between 2D maps in planar regions and 3D maps in elevation-rich environments via scene recognition using the Swift Scene Network (SwiftSceneNet). SwiftSceneNet is a lightweight hybrid Convolutional Neural Network (CNN)-transformer network developed as part of the ADAMS framework for semantic scene understanding. SwiftSceneNet integrates local geometric and global contextual cues with minimal inference latency, enabling reliable scene recognition for adaptive switching. ADAMS dynamically allocates computational resources and employs highfidelity 3D representations only when structural complexity requires them. Simulated and real-world tests confirm that SwiftSceneNet achieves an F-measure (F1-score) of 98.2% on a custom scene dataset and 96.5% on the Aerial Image Dataset for Emergency Response Applications version 2 (AIDERv2) at 25 frames per second (FPS). ADAMS yields a 32% reduction in Central Processing Unit (CPU) usage and a 65.7% reduction in memory consumption compared with Octree-based Mapping (OctoMap) and a 58.2% reduction relative to 3D Simultaneous Localization and Mapping (SLAM), while maintaining an Average Displacement Error (ADE) of 0.06 meters (m).

키워드