A Comparative Analysis of Visual SLAM Algorithms for Accurate Egomotion Estimation: Case Study with a Quadruped Robot Platform

Simultaneous Localization and Mapping (SLAM) algorithms are critical for enabling autonomous navigation in robotics. Specifically, the precision of the SLAM algorithm in accurately estimating the robot's position is crucial, as unreliable egomotion estimation can lead to significant failures or hazards. This thesis investigates the comparative performance of two leading SLAM algorithms, ORB-SLAM3 and RTAB-Map, by performing an experiment with a Unitree Go1 quadruped platform equipped with an Intel RealSense D435i camera in an industrial indoor environment. The motivation for this research arises from the need of precise pose estimation capabilities, especially in industrial settings where inaccuracies can result in costly damages. Moreover, existing literature presents relatively few comparisons of these two algorithms and often with contrasting results - suggesting that different environments or settings may influence SLAM algorithm performance significantly. This study aims to fill this gap and provide a method to benchmark SLAM algorithms in scenarios where motion capture systems are not available. As of our knowledge, this is the first experiment comparing RTAB-Map and ORB-SLAM3 in terms of pose estimation on a quadruped robot In the experiment, we conduct an accurate calibration of the visual and intertial sensors, employing the open-source Kalibr tool to calibrate the intrinsic and extrinsic parameters of the camera, and the Allan Variance method to model the inherent noise of the IMU system. We establish a reliable ground truth through widely available tools and the estimated poses are compared with the ground truth using Absolute Trajectory Error (ATE) and Relative Pose Error (RPE) metrics, alongside a qualitative analysis. The main goal of this study is to clarify the qualitative and quantitative distinctions between ORB-SLAM3 and RTAB-Map, emphasizing the specific considerations for employing SLAM algorithms on legged robots, and to also establish a methodological basis for future comparative studies on egomotion estimation accuracy.