Objective
The goal is to find one or more instances of one or more known objects in a scene captured by a 3D sensor.

Applications

  • Robot object grasping

  • Bin picking

  • Defect detection and quality checks

Challenges
The main difficulty of 3D registration is to guarantee fast cycle times while preserving robustness of the process. Indeed, in an industrial context where each mistake stops a whole production chain, reliabilities higher than 99% may be required. The high noise level of 3D sensors as well as the physical characteristics of industrial objects (no texture, specularities, …) make reaching the required robustness especially hard.

Proposed solution
In order to tacke the problem, CEA implemented an algorithmic chain based on a new local shape descriptor as well as a multi-dimensional pose analysis, with multiple patents pending. This algorithmic chain is GPU accelerated to guarantee optimal cycle times.
This solution is industrially deployed by our partner TRIDIMEO.

Current works
Current developments focus on mostly symmetric objects as well as integration with smart pluridigital robotic prehension.

Automatic detection of symmetry-breaking features

Computation time improvements relative to existing approaches for various point cloud sizes

timings of registration, small point cloud
timings of registration, medium point cloud
timings of registration, large point cloud
timings of ICP refinement