My research focuses on the processing of dense acquisitions of terrestrial LiDARs representing complex scenes.
With the evolution of acquisition devices, 3D point clouds are now an essential representation of real life scenes. The recent systems are able to acquire up to hundreds of millions of points in a single acquisition. If several acquisitions are needed to capture a large-scale scene, a whole building for instance, we obtain gigantic point clouds, i.e., composed of billions of points.
A local graph-based representation for gigantic point clouds
Our work tries to take advantage of the structure of the acquisitions, represented as depthmaps.
By constructing a graph from each acquisition and linking the constructed graphs together, it can be possible to process the underlying surface of gigantic point clouds, even on computers with limited memory.
This work tends to provide a solution to the simplification of point clouds in a "surface-aware" manner.
By selecting samples using a Poisson-disk sampling applied on a set of connected local graphs, the final samples exhibit good blue noise properties and a good spatial covering over the whole captured surface.
The versatility of the graph approach allows to simply model different metrics, thus generating different kind of samplings (uniform, curvature-aware).
This work is oriented towards the reconstruction of a surface mesh from a point cloud using a set of connected local graphs.
By computing a Centroidal Voronoi Tesselation of a graph, we show how a triangular mesh can be constructed from an acquisition.
Our current researchs now focus on the extension of this method to the reconstruction of point clouds obtained from a set of acquisitions.
Point cloud visualization and compression
During the beginning of my thesis, we worked on the problem of visualization and compression of point clouds.
By taking advantage of the lattice structure of depthmaps, we presented a multi-resolution structure for a point cloud based on a wavelet decomposition of the depthmaps. Moreover, we showed how a point cloud could be compressed using classical image compression schemes (JPEG2000, BPG, PNG) and how it compared to classical octree-based approaches.
Want to know more about me?
I am also available on the following social networks.