Diluka Moratuwage

Diluka Moratuwage

I am a lecturer at CQUniversity, based in Gladstone Marina Campus, Queensland, Australia.

Random Finite Set Based 3D Visual SLAM

Autonomous robots operating in GPS-denied environments require reliable localisation and mapping capabilities to navigate safely and build useful representations of their surroundings. This project developed a Random Finite Set (RFS) based 3D Visual SLAM framework that combines probabilistic state estimation with stereo vision to jointly estimate robot trajectory and construct three-dimensional maps.

The framework uses a particle filter to estimate the robot trajectory and a Gaussian Mixture Probability Hypothesis Density (GM-PHD) filter to estimate the map. Stereo visual odometry and triangulated 3D feature positions are used as inputs, while the resulting 3D map is visualised in ROS using OctoMap and RViz.

Research Motivation

SLAM in real-world environments is challenging because robots must estimate their own motion while simultaneously building a map from noisy, uncertain, and sometimes incomplete sensor observations. Random Finite Set theory provides a principled Bayesian framework for representing uncertainty in the number and state of environmental features, making it well suited to probabilistic SLAM and mapping problems.

Research Contributions

  • Developed an RFS-based framework for 3D Visual SLAM.
  • Integrated particle-filter-based robot trajectory estimation with GM-PHD map estimation.
  • Used stereo vision and visual odometry for localisation and 3D mapping.
  • Demonstrated the approach across handheld, aerial, and ground robotic platforms.
  • Implemented and visualised the system using ROS, RViz, and OctoMap.

Experimental Platforms

This work was demonstrated using several robotic and sensing platforms:

  • Handheld ZED stereo camera connected to a laptop.
  • ZED stereo camera mounted on an AscTec Pelican UAV.
  • Clearpath Husky ground robot fitted with a ZED stereo camera.

Demonstration Videos

Handheld ZED Stereo Camera Demonstration

Additional Handheld Demonstration

UAV Indoor Laboratory Demonstration

UAV Outdoor Campus Demonstration

Clearpath Husky Ground Robot Demonstration

Additional Husky Demonstration

Publications

  • (Add related journal/conference publications here.)

Future Directions

This work forms part of a broader research direction in probabilistic perception and localisation for autonomous robotic systems. Future extensions include tighter integration with multi-robot perception, semantic mapping, robust field deployment, and decision-making under uncertainty.