3d sensor

High-Resolution Three-Dimensional Sensing of Fast Deforming Objects: 

High-Resolution Three-Dimensional Sensing of Fast Deforming Objects Philip Fong Florian Buron Stanford University This work supported by:

Motivations: 

Motivations Considerable research on sensing 3D geometry of objects Focused on rigid objects and static scenes Moving and deforming objects found in many applications

Motivational Applications: 

Motivational Applications Navigation in dynamic environments Object recognition Human tissue modeling for surgical simulation and planning Robotic manipulation of fabric and rope

Goal: 

Goal Generate rangemaps (2.5D) from single images No temporal coherence assumption No restriction on motion Use commercially available hardware Minimize cost

Existing 3D Sensing Methods: 

Existing 3D Sensing Methods Time of flight Lidar, radar, sonar, shuttered light pulse Triangulation Laser stripe scanner Stereo Structured Light Can be implemented with commercial cameras and projectors Sinusoids / Moiré gratings (Takeda and Kitoh; Tang and Hung, Sansoni et al) Stripe patterns (Koninckx, Griesser, and Van Gool; Zhang, Curless, and Seitz; Caspi, Kiryati, and Shamir; Liu, Mu, and Fang)

Limitations of Existing Methods: 

Limitations of Existing Methods Requires scene be rigid and not moving Laser scanner Requires non-repeating texture Stereo vision Applied to deforming cloth (Pritchard and Heidrich) Requires known scene topology and known anchor points Sinusoids / Moiré gratings Requires multiple frames Restricts movement Spacetime Stereo (Davis, Ramamoothi, and Rusinkiewicz) Stripes In single frames, spatial resolution does not scale due to fixed number of encodings

System Overview: 

System Overview Idea: Combine colored stripes with sinusoid pattern Use sinusoidal pattern to produce dense rangemap Colored stripe transitions give sparse absolute depths Use to generate anchor points

System Geometry: 

System Geometry Camera projection center at (0,0,0) Projector at (px, py, pz) Parallel optical axes Pinhole projection model for camera and projector

System Overview: 

System Overview Image Demodulate Segment Phase Unwrap Find color transitions Label colors Label transitions Generate guesses Rangemap

Depth from Sinusoid: 

Depth from Sinusoid Projected sinusoid: Camera sees deformed sinusoid: Demodulate to get wrapped phase (Tang and Hung):

Segmenting: 

Segmenting Phase unwrapping assumes θ changes by less than 2π between pixels Segment image into regions based on phase variance (Ghiglia and Pritt) using snakes

Labeling Colors: 

Labeling Colors Label and score pixels with colors using Bayesian classifier PDFs of colors in hue space are approximated with a gaussian distribution

Labeling Color Transitions: 

Labeling Color Transitions Threshold change in hue between pixels along X direction Label each detected transition according to the projected pattern Based on color label scores in pixel windows to the left and right of transition Ignore transitions: Not consistent with projected pattern Over a max width

Generating Guesses: 

Generating Guesses In projected pattern Each transition appears only once Identifies unique vertical plane Intersect with ray corresponding to transition location in camera to compute depth Use as guesses in phase unwrapping

Phase Unwrapping: 

Phase Unwrapping Compute phase gradient assuming no jumps greater than 2π Integrate to get θu For each region compute k from median of the difference between guesses and θu Compute rangemap from θ

Results: Moving Speaker: 

Results: Moving Speaker 0.7mm (0.1%) RMS error compared to Cyberware 3030MS laser scanner

Results: Deforming Foam: 

Results: Deforming Foam Simple deformable object consisting of two types of foam System works in presence of color variation

Results: Flag Waving: 

Results: Flag Waving

Advantages: 

Advantages Spatial resolution scales with camera and projector resolution Temporal resolution scales with camera speed Not limited by projector speed Complex projector could be eliminated for static patterns

Limitations: 

Limitations Each segmented region must contain at least one recognized color transition Objects with many saturated colors are hard to sense Mitigated by choosing right set of colors in pattern Projected pattern must be bright enough to be seen Difficult to achieve outdoors

Conclusions / Results: 

Conclusions / Results Combined sinusoidal and colored stripe pattern is effective Produces good quality dense range maps of moving and deforming objects Works in presence of color variation Works in presence of fast motion and large deformations

Questions?: 

Questions? More at: http://www.stanford.edu/~fongpwf/research.html