Source-linked AI summary
Comparison of Various SLAM Systems for Mobile Robot in an Indoor Environment
Maksim Filipenko, Ilya Afanasyev
TL;DR
The paper addresses how different ROS-based SLAM systems perform for indoor mobile-robot localization and mapping across lidar, monocular-camera, and stereo-camera sensing. It evaluates them on a common sensor dataset from a prototype robot and reports accurate lidar results, strong visual systems, and sensor-specific limitations.
Problem
The paper examines comparative performance of ROS-based SLAM systems using different sensors for indoor mobile-robot localization and mapping.
Method
The authors collect synchronized 2D-lidar, monocular-camera, and stereo-camera data from a mobile-robot prototype and evaluate multiple SLAM systems using Absolute Trajectory Error.
Results
Cartographer and Hector SLAM achieve 0.024 m RMSE ATE, while RTAB map records 0.163 m and ORB SLAM 0.190 m RMSE ATE among visual systems.
Takeaways & Limitations
Lidar systems provide accurate localization and mapping, stereo systems provide metric localization without scaling modules, and monocular systems require additional scale recovery.
Takeaways & Limitations
Visual SLAM performance depends strongly on the mobile robot’s computational resources, whose hardware limitations remain for future study.
Abstract
from arXiv · showhide
This article presents a comparative analysis of a mobile robot trajectories computed by various ROS-based SLAM systems. For this reason we developed a prototype of a mobile robot with common sensors: 2D lidar, a monocular and ZED stereo cameras. Then we conducted experiments in a typical office environment and collected data from all sensors, running all tested SLAM systems based on the acquired dataset. We studied the following SLAM systems: (a) 2D lidar-based: GMapping, Hector SLAM, Cartographer; (b) monocular camera-based: Large Scale Direct monocular SLAM (LSD SLAM), ORB SLAM, Direct Sparse Odometry (DSO); and (c) stereo camera-based: ZEDfu, Real-Time Appearance-Based Mapping (RTAB map), ORB SLAM, Stereo Parallel Tracking and Mapping (S-PTAM). Since all SLAM methods were tested on the same dataset we compared results for different SLAM systems with appropriate metrics, demonstrating encouraging results for lidar-based Cartographer SLAM, Monocular ORB SLAM and Stereo RTAB Map methods.
I. INTRODUCTION
SLAM jointly builds maps and estimates robot location, using onboard sensors whose suitability affects indoor UGV operation. The paper compares ROS-based lidar, monocular, and stereo-camera systems.
- SLAM simultaneously builds a map of an unknown environment and estimates the robot’s location.Classical filter-based approaches use prediction and measurement-update steps, with EKF- and particle-filter implementations.
- Sensor choice is important for indoor UGV operation because inertial-only localization can accumulate significant navigation errors.The paper identifies lidar as a usual main sensor for indoor robot navigation and SLAM.
- Visual SLAM uses monocular or stereo cameras as low-cost passive sensors and includes feature-based and direct methods.The introduction lists representative systems including ORB-SLAM, LSD-SLAM, RTAB map, and DSO.
- The study investigates ROS implementations across three sensor classes: 2D lidar, monocular cameras, and stereo cameras.The tested systems are GMapping, Hector SLAM, Cartographer, LSD SLAM, ORB SLAM, DSO, ZEDfu, RTAB map, and S-PTAM.
II. RELATED WORK
Related studies compare ROS-based SLAM methods for indoor mobile-robot navigation, including lidar-only and mixed-sensor evaluations against trajectory references.
- Recent research compares various ROS-based SLAM methods for indoor navigation in mobile robotics.
- Some studies focus specifically on the efficiency of lidar-based SLAM approaches.One cited comparison evaluates Gmapping, KartoSLAM, Hector SLAM, CoreSLAM, and LagoSLAM.
- A prior mixed-sensor study compared trajectories from monocular and stereo cameras, lidar, and Kinect 2 in a homogeneous indoor environment.Its systems included monocular ORB-SLAM and DPPTAM, stereo ZedFu, and Kinect-related RTAB map packages.
III. SYSTEM SETUP
The study uses a Labcar unmanned ground vehicle prototype equipped with lidar and monocular and stereo cameras for indoor dataset acquisition and SLAM evaluation.
- The experimental platform is a Labcar unmanned ground vehicle equipped with a computational module, 2D lidar, monocular camera, and ZED stereo camera.
- The prototype combines lidar and both camera types so the same platform can supply data for multiple SLAM sensor classes.
- The dataset acquisition platform is based on a Traxxas 7407 radio-controlled car model.A ground station is used for execution of the recorded sensor data.
B. Robot model
The Labcar is integrated into ROS for consistent robot modeling and coordinate handling across SLAM experiments, with its hardware and software specified in platform tables.
- All tested SLAM systems integrate with ROS either fully or through wrappers.
- URDF and tf transfer UGV position data into a common coordinate frame for more consistent cross-system comparison.The common representation also reduces human errors during experiments.
- The robot model is represented in a ROS/Gazebo simulator environment.
IV. METHODOLOGY OF THE EXPERIMENT
The experiment used a teleoperated mobile robot in a typical indoor office environment, with the trajectory perimeter marked for reference. Sensor data and robot-state visualizations were collected in ROS for offline SLAM processing.
- The mobile robot followed a teleoperated closed-loop trajectory around a rectangular work area in a typical office environment.
- A thread marked the UGV trajectory perimeter and served as ground truth, although the robot coincided with it only during direct movement.Turns between straight segments had an approximately 1 m radius.
- Sensor data were collected in ROS bags, and the ground station ran SLAM algorithms on the recorded data.
- RViz visualized the robot model, forward-camera view, 2D lidar point cloud, and built map for the test dataset.
B. Metrics for evaluation
The evaluation compared each tested system’s trajectory with a Hector SLAM reference trajectory using Absolute Trajectory Error and summary statistics.
- Absolute Trajectory Error compared each tested system’s trajectory with more accurate UGV trajectory data.Hector SLAM data were used to verify visual SLAM systems under the assumption that Hector SLAM was more accurate.
- The reference coordinates were defined at times t_i in the Hector SLAM coordinate frame, while the exploring method supplied corresponding trajectory data.
- The Hector SLAM trajectory was interpolated with a first-degree polynomial before absolute trajectory error was computed.
- Evaluation used RMSE, mean, median, standard deviation, minimum, and maximum statistics.
A. Odometry analysis
Odometry analysis evaluated localization trajectories with Absolute Trajectory Error metrics and identified RTAB map as strongest by RMSE, followed by monocular ORB SLAM.
- Table V presents the Absolute Trajectory Error metrics used to compare the different systems.
- RTAB map was among the best tested methods for mobile-robot localization in terms of RMSE.
- Monocular ORB SLAM was estimated as the second most accurate method, while ZEDfu produced the worst results in these experiments.
B. Maps analysis
Map analysis compared maps generated by lidar and visual SLAM systems. Cartographer was recommended for 2D lidar SLAM, while visual methods differed in map density and navigation suitability.
- Figure 6 presents maps generated by the various SLAM systems.
- Cartographer was recommended as the best 2D lidar SLAM choice because global optimization and local probabilistic updates improve robustness to environmental changes.
- GMapping produced an inaccurate map, whereas Hector SLAM and Cartographer produced quite similar maps.
- LSD SLAM and DSO produced dense point-cloud maps supporting 3D scene recovery and object detection.
- ORB SLAM, RTAB map, and S-PTAM produced sparse point-cloud maps that are suitable for navigation but make 3D scene recovery difficult.
VI. CONCLUSION
The study compares ROS-based SLAM systems using a common mobile-robot dataset and reports distinct strengths across lidar, monocular, and stereo approaches. Cartographer and Hector SLAM were accurate for lidar, while RTAB-Map and ORB SLAM were notable among visual systems, with scale and computational constraints remaining important boundaries.
- Experimental comparison: The experiments processed shared sensor data from 2D lidar, monocular, and stereo cameras using multiple ROS-based SLAM methods and metric comparisons.The robot operated in a typical office environment, and the collected dataset was processed offline.
- 2D lidar systems: Hector SLAM and Cartographer provided accurate lidar-based localization and map building, with both achieving an ATE RMSE of 0.024 m.Both trajectories coincided with the marked line on the floor.
- 2D lidar systems: Cartographer was more robust to environmental changes because it combines global map optimization with local probabilistic map updates.
- Monocular systems: Monocular LSD SLAM, ORB SLAM, and DSO could solve localization only with an additional module for scale recovery, while other monocular systems lost track because of insufficient features.
- Stereo systems: Stereo systems provided metric localization without additional scaling modules and also built 3D metric point clouds.
- Visual-system comparison: RTAB-Map achieved the best visual localization result with ATE RMSE of 0.163 m, whereas ORB SLAM was the most robust and stable at 0.190 m.RTAB-Map could lose track near monochrome walls.
- Scope and future work: Visual SLAM performance depends strongly on the robot's computational resources, motivating future study of hardware limitations and their effects on efficiency.