8/14/2019 Planar Objects and Articulation Models

1/17

SA-1

Marker-less Object Perception

and Articulation Discovery

Jrgen SturmKurt Konolige

8/14/2019 Planar Objects and Articulation Models

2/17

Previous work

Learn articulation models from pose observations Model selection

Rotational model

Prismatic model

Non-parametric LLE/GP model Structure discovery

8/14/2019 Planar Objects and Articulation Models

3/17

Microwave Door: Observations

microwave pose observationsfrom motion capturing studio

8/14/2019 Planar Objects and Articulation Models

4/17

M crowave Door: LearneModel

learned model for microwave door

8/14/2019 Planar Objects and Articulation Models

5/17

Cabinet with Two Drawers

learned models and structure

8/14/2019 Planar Objects and Articulation Models

6/17

Research question

Can we get rid of artificial markers for poseregistration?

Can we learn articulation models in unpreparedenvironments?

8/14/2019 Planar Objects and Articulation Models

7/17

Choosing the sensor

Use stereo vision? Videre stereo camera

Projected light

8/14/2019 Planar Objects and Articulation Models

8/17

Stereo vision + structured light

Structured lightprojector adds much

texture to scene Disparity image is

dense

Dense depth video

8/14/2019 Planar Objects and Articulation Models

9/17

Problem formulation

Dense stereo data Objects have rectangular shape

Unknown position

Unknown size

Unknown orientation

8/14/2019 Planar Objects and Articulation Models

10/17

First approach

Segment planes

Search for edges (Canny)

Search for lines (Hough)

Line intersections corner candidates

Find width, height

Optimize fit on distance

transform (chamfermatching)

8/14/2019 Planar Objects and Articulation Models

11/17

First approach

Segment planes

Search for edges (Canny)

Search for lines (Hough)

Line intersections corner candidates

Find width, height

Optimize fit on distance

transform (chamfermatching)

XDepends on good edge visibility

Poor performance on doors

Way too complicated!

8/14/2019 Planar Objects and Articulation Models

12/17

Second approach

Segment planes Pick random seed pixel

Iteratively optimize insmall steps

width (from left) width (from right)

height (from bottom)

height (from top)

rotation

Objective function fill ratio of rectangle

slight bias term thatfavors larger objects

8/14/2019 Planar Objects and Articulation Models

13/17

More examples

Cabinet door Cabinet drawer

Fuse door

Book

Carton

8/14/2019 Planar Objects and Articulation Models

14/17

Object Tracking

Track observations overtime

Noise

Partial observations

Ambiguities Front/backside flips

Rotations of90/180/270deg

Track assignment

Data association

8/14/2019 Planar Objects and Articulation Models

15/17

Discover articulated objects

Learn articulation modelsfor tracks

Measure model fit

Estimate currentobject configuration

Make pose predictions forunseen configurations

8/14/2019 Planar Objects and Articulation Models

16/17

Conclusions

simple object detection full pose estimates

articulation model learning on natural features ispossible

(currently) limited to rectangular shaped objects

implemented as ROS package planar_objects box_detector

box_tracker articulation_learner

Demo after this talk in green room

8/14/2019 Planar Objects and Articulation Models

17/17

Future work

ground truth evaluation improve objective function (use occ/free/unknown)

appearance-based matching

add rotational articulation model

improve plane extraction using surface normals optimize code (currently 1-4s per frame)

ICRA paper