Abstract. This work introduces a novel solution for localizing objects based on search strings and freely available Google SketchUp models. To this end we automatically download and preprocess a collection of 3D models to obtain equivalent point clouds. The outdoor scan is segmented into individual objects, which are sequentially matched with the models by a variant of iterative closest points algorithm using seven degrees of freedom and resulting in a highly precise pose estimation of the object. An error function evaluates the similarity level. The approach is verified using various segmented cars and their corresponding 3D models.
Ride the Storm: Navigating Through Unstable Periods / Katerina Rudko (Belka G...
Object Indentification Using 3D SketchUp Models in Environment Scans
1. Object Identification Using
3D SketchUp Models in
Environment Scans
Flavia Grosan, Alexandru Tandrau
Prof. Dr. Andreas Nüchter
Thursday, May 12, 2011
2. Introduction
n “I can’t find my Audi A4. Bot, please find it for me!”
Thursday, May 12, 2011
3. Introduction
n SLAM
n Laser range scanners
n ICP
n Semantics
Thursday, May 12, 2011
4. State of the Art
n Horn - closed form solution for ICP
n translation, rotation, scale
n Nüchter - semantic mapping
n determine coarse features: walls, floors
n trained classifier to identify more delicate objects
Thursday, May 12, 2011
5. State of the Art
n Object Localization
n Li-Jia - 2D object localization
n Meger - Semantic Robot Challenge
n Kestler - probabilistic representation
n manual labeling needed
n maintains internal neural net trained data
n Lai and Fox - Google Warehouse to train classifiers
n Albrecht - CAD models and ontologies
Thursday, May 12, 2011
6. Scientific Contribution
n Combine laser scanning with object detection and
localization
n Simple scan matching instead of classifiers and
probabilistic approaches
n Evaluates Google 3D Warehouse - a new, large 3D
model database
Thursday, May 12, 2011
7. From Model to Point Cloud
n Google 3D Warehouse - collection of
user made SketchUp models
n A model is composed of:
n Faces
n ComponentInstances
n Groups
Thursday, May 12, 2011
8. From Model to Point Cloud
n Additional sampling procedure needed
n Add random points inside each
triangular face proportionally to its area
n Center the point cloud around its
centroid and bound the coordinates in
[-α, α]
Thursday, May 12, 2011
9. Model - Scan Matching
Is the model present in the scan? If so, where?
Thursday, May 12, 2011
10. Model - Scan Matching
n Ground Removal
n Stiene et al.
n Compute gradient
between closest
points in the same
vertical sweep plane
n A point is classified as
ground if -θ ≤ αi,j ≤ θ
Thursday, May 12, 2011
11. Model - Scan Matching
Object segmentation by region growing
with starting point
Thursday, May 12, 2011
12. Model - Scan Matching
Object segmentation by region growing
with starting point
Thursday, May 12, 2011
13. Model - Scan Matching
n The centroid of the segmented object is the new
system origin. The object coordinates are bounded
in [-α, α] (center and scale step)
n Modified ICP (SICP) to match model and scan
n Scale
n Favor rotations on the y-axis (wheels on the ground)
n Points are linked in both directions (scan to model,
model to scan)
n Recover transformation matrix to original scan
Thursday, May 12, 2011
14. Model - Scan Matching
SICP animation
Thursday, May 12, 2011
15. Model - Scan Matching
SICP animation
Thursday, May 12, 2011
16. Model - Scan Matching
n S - the scan points, M - the model points
n c(p) ∈ M - the model point which is closest to p ∈ S
the error function penalizes points in the scan
which have no correspondence in the model
Thursday, May 12, 2011
17. Experiments and Results
n acquired scans using a Riegl VZ-400 3D laser
scanner in the Jacobs University parking lot
n segmented 5 different cars based on starting points
n automatically downloaded relevant Google
SketchUp models and pre-processed them
(resampling, scale & center)
n SICP with 4 starting rotations around the vertical
axis
Thursday, May 12, 2011
18. Mercedes C350
n 8920 points in scan
n 89 models available in Google Warehouse
Thursday, May 12, 2011
26. Conclusions
n Segmented objects above 10,000 points behaved well in SICP
n controlled in practice by taking more scans Perfect Match %
n Number of Google Warehouse models 0 10 20 30 40
n Volkswagen Golf - 200+ models Audi A4
n Citroen C5 - 11 models VW Golf
Mercedes C350
n ~ 80 models needed to get good matches
Thursday, May 12, 2011
27. Conclusions
n Identified Volkswagen Golf - an older variant
Best matches distribution VW Golf
n SICP ranks higher Google models
resembling the old Golf version
Newer Golf
37% Older Golf
63%
n SICP identifies similar shapes
n Mercedes C350 vs. other car brands
Thursday, May 12, 2011
28. Conclusions
n SICP solves the goal finding problem
n Automatic scan segmentation
n Correct identification of Audi A4
n Next 2 matches - also cars, different brands
Thursday, May 12, 2011
29. Future Work
n Refine model search in Google Warehouse
n Improve the error function
n Tackle indoor scenarios
n SICP - extendable to full-scene understanding
n Create an online platform for SICP
n Integrate SICP as a plugin for RoboEarth
Thursday, May 12, 2011