This project showcases a complete implementation of visual-inertial SLAM using the Intel RealSense D455 and RTAB-Map within the ROS environment.

The project highlights practical integration through the following key capabilities:

• Real-time 3D mapping using RGB-D and IMU data
• Visual odometry with both ORB and SIFT feature detectors (compared side-by-side)
• Loop closure detection to correct trajectory drift
• Pose graph optimization for global consistency
• Generation of dense point cloud maps
• Visualizations of the trajectory, graph structure, and loop closures

This project was implemented on a mobile robotic system for testing visual-inertial SLAM performance in real-world indoor settings.

• Intel RealSense D455
  • RGB-D camera with onboard IMU
  • Used for both visual odometry and depth mapping
• Clearpath Husky UGV
  • Served as the mobile base to replicate realistic ground robot operation
• Onboard Laptop
  • Runs ROS Noetic and all mapping processes
  • Specs: NVIDIA GTX 1660 Ti (mobile), Intel Core i7-9750h, 16GB RAM

• Mapping conducted in structured indoor spaces
• Included hallways and rooms with furniture and varying layout
• Designed to test loop closures, graph optimization, and keypoint robustness

Structured Hallway

Lab Space with Dense Obstacles

The SLAM system uses RTAB-Map to incrementally build 3D maps while tracking the robot's position over time. It combines visual odometry, loop closure, and graph optimization to maintain an accurate and drift-corrected trajectory. As the robot explores, the system detects previously visited locations and refines the map for global consistency.

Loop closures were automatically detected by comparing current visual data with previously seen scenes. When a match was found, RTAB-Map estimated the correction and updated the pose graph to align the trajectory. This helped reduce drift and improve map consistency during longer sessions.

Loop Closure Detection

Feature Matching Camera odometry frame and 3D point cloud projection

In the example above, a loop closure was detected between Frame 135 and Frame 87, applying a correction of Δx = -0.0117 m and Δθ = -0.43° to the trajectory.

After detecting loop closures, RTAB-Map applies graph optimization to refine the map and reduce accumulated drift.
This process removes redundant poses, merges overlapping data, and ensures the trajectory remains globally consistent.

Pose Graph Before Optimization

Pose Graph After Optimization

The global graph was reduced from 445 nodes to 407 nodes (an 8.5% reduction).
This optimization improved the alignment of the robot’s trajectory, removing drift and enhancing overall map accuracy.

Feature detection was tested with both ORB and SIFT to evaluate the trade-off between speed and accuracy.
ORB is fast and lightweight, making it suitable for real-time operation, while SIFT is more computationally demanding but detects a larger number of robust features, especially in low-texture areas.

ORB Feature Detection

SIFT Feature Detection

• ORB – Efficient, real-time performance but fewer detected keypoints.
• SIFT – Higher accuracy in challenging environments but slower processing.

This comparison highlights the balance between real-time operation and feature richness in SLAM implementations.

Below are two examples of 3D point cloud maps generated using RTAB-Map with the Intel RealSense D455. These maps were built in real time while maintaining a drift-corrected trajectory through loop closure and graph optimization.

Structured Hallway

Lab Space with Dense Obstacles

• Lab Environment – Captured in a confined indoor lab with dense obstacles such as tables and chairs. The map clearly shows the workspace layout, with a well-defined trajectory loop in the center.
• Corridor Environment – Captured along a long structured hallway with multiple turns. The map maintains wall alignment and straight trajectories despite extended travel distance.

These results highlight the system’s ability to adapt to both dense, cluttered spaces and open, structured layouts, producing consistent maps in varied indoor settings.