Détection et localisation d'objets 3D par apprentissage profond en topologie capteur

Abstract : This work proposes a novel approach for detection and localisation of objects in 3D LiDAR scenes aquired via Mobile Mapping Systems. While this task is often treated on a voxel grid representations of the point cloud, our method offers to use the point cloud in sensor topology, thus avoiding a discretisation step. This representation of the point cloud is used as an input for a CNN that extracts 3D positions and dimensions of objects in the scene. As far objects in the scene tends to be mixed with the background when seen in the sensor topology, we offer to enhance the 3D detection by fusing the 3D predictions with 2D object detections performed on optical images.
Document type :
Preprints, Working Papers, ...
Complete list of metadatas

https://hal.archives-ouvertes.fr/hal-02100719
Contributor : Pierre Biasutti <>
Submitted on : Tuesday, April 16, 2019 - 10:57:43 AM
Last modification on : Friday, April 19, 2019 - 1:46:50 AM

File

GRETSI_2019 (2).pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-02100719, version 1

Collections

Citation

Pierre Biasutti, Aurélie Bugeau, Jean-François Aujol, Mathieu Brédif. Détection et localisation d'objets 3D par apprentissage profond en topologie capteur. 2019. ⟨hal-02100719⟩

Share

Metrics

Record views

13

Files downloads

46