SGBA: Semantic Gaussian Mixture Model-Based LiDAR Bundle Adjustment

This repository implements SGBA, a LiDAR bundle adjustment system that uses semantic Gaussian mixture models for robust pose estimation.

Quick Start

Prerequisites

• ROS (tested with ROS Noetic)
• PCL
• Eigen3
• OpenMP

Build

cd ~/catkin_ws
catkin build semantic_ba

Configuration

Edit config/read_label.yaml to set the test mode and parameters.

Running Tests

1. Virtual Planes Test

Test with synthetic planar data for algorithm validation.

Setup:

1. Set test mode in config/read_label.yaml:

   test_mode: "virtual_planes"

2. Configure virtual plane parameters:

   virtual_planes:
     pcd_file: "/path/to/your/virtual_plane_data.pcd"
     max_frames: 20
     use_identity_transforms: true

Run:

rosrun semanticBA algoTest

2. KITTI Semantic Test

Test with real KITTI semantic dataset.

Setup:

1. Download KITTI semantic dataset

2. Set test mode in config/read_label.yaml:

   test_mode: "kitti"

3. Configure KITTI parameters:

   semantic_kitti:
     kittiPath: "/path/to/kitti_semantic/sequence"
     frameCnt: 400      # starting frame
     winSize: 20        # sliding window size
     label2Read: [80, 71, 40, 48]  # semantic labels to use

Run:

rosrun semanticBA algoTest

Controls

• Press 'n' to proceed to next iteration
• Press 's' to show performance statistics
• Press 'q' to quit

Dataset Format

• Virtual Planes: PCD files with intensity (frame ID) and curvature (virtual semantic label)
• KITTI: Standard KITTI semantic format with point-wise semantic labels

Video: