SLAM Based Robot Navigation Using ORB Slam3 — Implementation & Analysis

Abstract — SLAM Based Robot Navigation Using ORB Slam3

Abstract This study presents an enhanced visual SLAM (Simultaneous Localization and Mapping) framework that integrates ORB-SLAM3 with the YOLOv5 real-time object detection model to improve pose accuracy in dynamic environments. Although ORB-SLAM3 achieves robust performance in static scenes, its reliance on ORB feature tracking often degrades accuracy in the presence of moving objects. To overcome this limitation, YOLOv5 is employed to identify dynamic regions in each video frame, enabling the system to remove motion-related feature points before matching. This filtering mechanism reduces the influence of dynamic objects on trajectory estimation and enhances overall system robustness. The proposed method was evaluated using dynamic datasets, including BONN and TUM RGB-D, and further validated through real-world experiments with an Intel RealSense D435i camera. Experimental results demonstrate substantial improvements in pose accuracy compared with the baseline ORB-SLAM3 and the RTAB-Map system, confirming the effectiveness of the YOLOv5-assisted ORB-SLAM3 integration in dynamic scenes.

Operational Methodology & Framework

Keywords: ORB-SLAM3, YOLOv5, dynamic environments, pose estimation, visual SLAM 1. Introduction The ability of a mobile robot to autonomously navigate an unknown environment relies heavily on its capacity to simultaneously localize itself and construct a consistent map of its surroundings. This problem—commonly referred to as Simultaneous Localization and Mapping (SLAM)—is central to mobile robotics and has been extensively studied over the past two decades. Classical SLAM techniques, such as Extended Kalman Filter SLAM (EKF-SLAM) , FastSLAM , and GraphSLAM , have proven effective in static environments, where the assumption of scene stability holds. However, these systems tend to degrade significantly in dynamic settings, where moving objects introduce incorrect feature associations, partial occlusions, and temporal inconsistencies in the map. As noted in a recent comprehensive survey by Chen et al. , conventional SLAM approaches struggle in unstructured or non-rigid scenes, prompting a shift toward more adaptive solutions.

Experimental Findings & Validation

Visual SLAM (vSLAM), which relies on camera inputs rather than LiDAR or other expensive sensors, has emerged as a promising approach for navigation in both indoor and outdoor environments. Algorithms such as ORB-SLAM2 and its successor ORB-SLAM3 are widely adopted due to their robustness, support for multiple sensor modalities, and open- source implementations. Nevertheless, both systems assume that the environment remains largely static. In real-world scenarios—particularly in crowded or dynamic indoor environments—this assumption is often violated, leading to degraded pose estimation and corrupted mapping due to the inclusion of transient, non-static features.

Scope & Practical Applications

• Real-time automated control and signal telemetry
• Embedded verification testbeds for institutional research
• High-throughput data streaming and anomaly alerting
• Performance optimization and resource consumption minimization
• Full integration ready for IEEE dissertation submission