Dense 3D SLAM System with Point-based Fusion
Embedded Files
Dense Simultaneous Localization and Mapping (Dense SLAM) is a powerful technique in computer vision and robotics, enabling a system to estimate its trajectory and build a detailed, per-pixel 3D map of unknown environments using visual sensor data—most notably from RGB-D cameras. Unlike sparse SLAM methods, which track only a small set of discrete feature points, dense SLAM captures and fuses information from the entire depth image, producing highly accurate and finely detailed reconstructions of the surrounding scene. This comprehensive mapping capability is vital for applications such as autonomous robot navigation, augmented and virtual reality, and 3D scene reconstruction, where detailed geometry and high accuracy are required for reliable interaction and decision-making
The dense SLAM pipeline developed for this project integrates two core techniques: Iterative Closest Point (ICP) for estimating the camera's motion between frames and point-based fusion for building the global 3D map. ICP aligns the current RGB-D frame with the existing map by minimizing the point-to-plane distances of matched points and efficiently refines pose estimates to achieve robust tracking. With each new frame, the point-based fusion method updates or merges observed 3D points into a weighted point cloud, resulting in a memory-efficient yet accurate representation of the scene. My primary contributions include developing the fusion pipeline, fine-tuning ICP parameters, and validating the system through quantitative and qualitative evaluation.
The dense SLAM pipeline developed for this project integrates two core techniques: Iterative Closest Point (ICP) for estimating the camera's motion between frames and point-based fusion for building the global 3D map. ICP aligns the current RGB-D frame with the existing map by minimizing the point-to-plane distances of matched points and efficiently refines pose estimates to achieve robust tracking. With each new frame, the point-based fusion method updates or merges observed 3D points into a weighted point cloud, resulting in a memory-efficient yet accurate representation of the scene. My primary contributions include developing the fusion pipeline, fine-tuning ICP parameters, and validating the system through quantitative and qualitative evaluation.
The system was evaluated on the ICL-NUIM synthetic dataset, a widely used benchmark that provides both ground truth poses and 3D models for thorough quantitative assessment
The system was evaluated on the ICL-NUIM synthetic dataset, a widely used benchmark that provides both ground truth poses and 3D models for thorough quantitative assessment
In summary, the project successfully implemented a dense 3D SLAM system capable of:
Achieving alignment errors under 6cm, consistently converging within 10 iterations.
Producing a dense 3D map from 200 RGB-D frames.
Reaching a compression ratio of 8.35%, effectively minimizing memory usage while retaining accurate scene geometry.
Page updated
Google Sites
Report abuse