Lightweight Visual SLAM Optimized with YOLOv11 and Its Applications

Abstract

Visual Simultaneous Localization and Mapping (vSLAM) remains computationally challenging for resource-constrained platforms, particularly when integrating robust semantic understanding. This paper presents a lightweight optimization framework for vSLAM that leverages the efficiency of YOLOv11 for real-time object-level semantic segmentation. Our approach strategically embeds YOLOv11's object detection output into the ORB-SLAM3 pipeline to enable dynamic feature culling and semantic-aided loop closure, significantly reducing the computational load associated with processing redundant visual features in non-informative image regions. By constructing a transient semantic map, the system prioritizes feature extraction and matching on structurally significant and semantically stable objects, enhancing both tracking accuracy and mapping utility while minimizing processing latency. Extensive evaluations on public datasets (e.g., TUM RGB-D, KITTI) demonstrate that our optimized system reduces average pose tracking error by 18% and decreases CPU utilization by over 32% compared to standard ORB-SLAM3, all while maintaining real-time performance on an embedded Jetson AGX Orin platform. The practical efficacy of the system is further validated through two application case studies: enhanced AR navigation in dynamic indoor environments and precise payload localization for an agricultural inspection drone. This work establishes a viable pathway for deploying intelligent, semantics-aware vSLAM on edge devices, effectively balancing accuracy, efficiency, and contextual awareness.