KinectFusion - Real-Time depth image fusion and dense surface mesh extraction


Abstract

Implementation of the KinectFusion approach to generating three-dimensional models from depth image scans. Here, the original method has been extended with the MarchingCubes algorithm to allow exporting the model as a dense surface mesh. Realized in modern C++14 and CUDA to allow real-time reconstruction. Developed in the context of an interdisciplinary project in cooperation with the Chair for Computer Aided Medical Procedures and Augmented Reality and Dynamify. Feel free to have a look at the library source and a sample application.

Readme

This is an implementation of KinectFusion, based on Newcombe, Richard A., et al. KinectFusion: Real-time dense surface mapping and tracking. It makes heavy use of graphics hardware and thus allows for real-time fusion of depth image scans. Furthermore, exporting of the resulting fused volume is possible either as a pointcloud or a dense surface mesh.

Features

  • Real-time fusion of depth scans and corresponding RGB color images
  • Easy to use, modern C++14 interface
  • Export of the resulting volume as pointcloud
  • Export also as dense surface mesh using the MarchingCubes algorithm
  • Functions for easy export of pointclouds and meshes into the PLY file format
  • Retrieval of calculated camera poses for further processing

Dependencies

  • GCC 5 as higher versions do not work with current nvcc (as of 2017).
  • CUDA 8.0. In order to provide real-time reconstruction, this library relies on graphics hardware. Running it exclusively on the CPU is not possible.
  • OpenCV 3.0 or higher. This library heavily depends on the GPU features of OpenCV that have been refactored in the 3.0 release. Therefore, OpenCV 2 is not supported.
  • Eigen3 for efficient matrix and vector operations.

Prerequisites

  • Adjust CUDA architecture: Set the CUDA architecture version to that of your graphics hardware SET(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-O3 -gencode arch=compute_52,code=sm_52) Tested with a nVidia GeForce 970, compute capability 5.2, Maxwell architecture
  • Set custom opencv path (if necessary): SET("OpenCV_DIR" "/opt/opencv/usr/local/share/OpenCV")

Usage

#include <kinectfusion.h>

// Define the data source
XtionCamera camera {};

// Get a global configuration (comes with default values) and adjust some parameters
kinectfusion::GlobalConfiguration configuration;
configuration.voxel_scale = 2.f;
configuration.init_depth = 700.f;
configuration.distance_threshold = 10.f;
configuration.angle_threshold = 20.f;

// Create a KinectFusion pipeline with the camera intrinsics and the global configuration
kinectfusion::Pipeline pipeline { camera.get_parameters(), configuration };

// Then, just loop over the incoming frames
while ( !end ) {
    // 1) Grab a frame from the data source
    InputFrame frame = camera.grab_frame();

    // 2) Have the pipeline fuse it into the global volume
    bool success = pipeline.process_frame(frame.depth_map, frame.color_map);
    if (!success)
        std::cout << "Frame could not be processed" << std::endl;
}

// Retrieve camera poses
auto poses = pipeline.get_poses();

// Export surface mesh
auto mesh = pipeline.extract_mesh();
kinectfusion::export_ply("data/mesh.ply", mesh);

// Export pointcloud
auto pointcloud = pipeline.extract_pointcloud();
kinectfusion::export_ply("data/pointcloud.ply", pointcloud);

For a more in-depth example and implementations of the data sources, have a look at the KinectFusionApp.

License

This library is licensed under MIT.