3d imaging with tof camera - khu.ac.krcvlab.khu.ac.kr/cvlecture24.pdf · tof-to-left occlusions:...

48
3D Imaging with ToF Camera

Upload: others

Post on 25-May-2020

20 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

3D Imaging with ToF Camera

Page 2: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Time-of-Flight Principle

Reflected IR shows phase delay proportional to the distance

from the camera.

Time-of-flight of Light Distance

: It is not simple to measure the flight time directly at each pixel

of any existing image sensor

Page 3: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Phase Delay Measurement

Q1 through Q4 are the amount of electrons measured at each

corresponding time.

In real situations, it is difficult to sense electric charge at certain

time instance

)(dt

Page 4: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Phase Delay Measurement

Distance

21

43arctan2

)(2 QQ

QQcdt

c

21

43

21

43 arctan2

arctan2 qq

qqc

qq

qqc

Assumption: Single reflected IR signal

In principle, amplitude of the reflected IR does not affect the depth

calculation.

Page 5: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Multiple IR Signals

- Large Sensor Pixel

- Scattering

- Multipath

- Motion Blur

- Transparent Object

In real situations, multiple reflected IR signals with different phase

delays & amplitudes can be superposed.

)()(

)()(arctan

2)(

2211

4433

qqqq

qqqqcdt

We do not know how many IR signals will be superposed.

Page 6: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Large Sensor Pixel

In order to increase sensitivity,

- large pixel size or pixel binning

IR signal #1

IR signal #2

)()(

)()(arctan

2)(

2211

4433

qqqq

qqqqcdt

Page 7: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Multiple light reflections between the lens and the sensor

Light Scattering

Light scattering [1]

[1] “Real-time scattering compensation for time-of-flight camera”, CVS07

Page 8: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Multipath Errors

IR LED

Sensor

Multipath Interference Depth error in concave objects

)()(

)()(arctan

2)(

2211

4433

qqqq

qqqqcdt

Page 9: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Motion Blur

Moving camera/object within single integration time make wrong

depth calculation

Image sensor

Moving Object Moving Object

Page 10: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Motion Blur

The characteristic of Tof motion blur is different from color

Overshoot Blur

Undershoot Blur

Overshoot Blur

Page 11: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Motion Blur

We use a set of cycles for depth calculation

In motion blur case, multiple IR signals come in sequentially

Reflected IR #1 & #2

TimeInteg.

1Q

2Q

3Q

4Q

Emitted IR

))1(())1((

))1(())1((arctan

2)(

2211

4433

qnnqqnnq

qnnqqnnqcdt

Page 12: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

ToF Deblurring

Blur Detection

Blur Level

Input Depth

Deblurred Depth

- There are some relations between Q1~Q4 1Q

2Q

3Q

4Q

4321 QQQQ KQQQQ 4321

- We assume 2-Layer blur case

: single flat foreground + single flat background

Page 13: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Transparent Object

2-Layer approximation of transparent object

))1(())1((

))1(())1((arctan

2)(

2211

4433

qqqq

qqqqcdt

- Sometimes 2-Layer is not enough

- Multiple reflection between objects (when they are close)

- In most cases, they have specular surface

Page 14: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Transparent Object

)(

)(arctan

21

43

QQ

QQtd

)ˆˆ()(

)ˆˆ()(arctan

2121

4343

QQQQ

QQQQtd

Depth

IR-Intensity

Transparent object

Now, amplitude matters

Page 15: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Transparent Object

)ˆˆ()(

)ˆˆ()(arctan

2121

4343

QQQQ

QQQQtd

Page 16: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Transparent Object

Page 17: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Due to the variation of the number of collected electrons during the integration time the repeatability of each depth point varies

Integration time-related Error

Integration Time: 30(ms) Integration Time: 80(ms)

Page 18: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Due to the non-uniformity of IR illumination and reflectivity variation of objects use a polynomial fitting model

Amplitude-related Errors

Amplitude image of a planar object

with a ramp image. Parts of the ramp

are selected for calibration (blue

rectangle).

The depth samples (blue) and the

fitted model (green) to the error

x(pixel)

y(p

ixe

l)

Amplitude

Err

or(

m)

0

0.001

0.003

0.004

0.002

0

1 0.5

Page 19: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Light attenuates according to the law of inverse square

Amplitude Correction

Distance-based intensity correction [18]

Page 20: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Kinect Principle (1/3)

Basically, it is based on structured light principle

IR Speckle

Pattern

Page 21: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Kinect Principle (2/3)

0. Calibrate source and

detector

1. Known IR pattern is

projected from the source

2. Detector identify each

dot (or set of dots)

3. Triangulate to calculate

depth

Page 22: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Kinect Principle (3/3)

- Random speckles identify x,y locations

- Orientation and shape of the speckles change along distance

identify z location

Page 23: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

ToF vs Kinect

Kinect Fusion SAIT & KAIST using ToF Camera

3D Reconstruction using multiple depth images

Page 24: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

24

Depth/Point Cloud Processing

3D Features

3D Filtering

Registration Surface Processing

Page 25: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Distortion Upon Materials

• Conventional approaches assume the Lambertian materials.

• Various surface materials exhibit the complex light interaction, causing the non-linear distortion on light transport.

• Depth cameras suffer from the depth distortion upon material properties.

• The type of distortion varies upon the sensing principle of depth cameras.

25

Page 26: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Cameras

• We provide the distortion analysis based on two sensor types: A Time-of-Flight and a structured light sensor

[Swissranger] [Kinect]

26

Page 27: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Distortion – Lambertian

• Material affects the sensing performance (Lambertian)

All existing 3D sensing

techniques are limited to

Lambertian object. Sensor

IR LED

Sensor

Projector

ToF depth camera Structured light depth camera

27

Page 28: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Distortion – Specularity

• Non-Lambertian materials causes the failure in sensing reflected signal (Specularity)

Sensor

IR LED

Sensor

Projector

ToF depth camera Structured light depth camera

28

Page 29: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Distortion – Translucency

• Non-Lambertian materials causes the failure in sensing reflected signal (Translucency)

Sensor

Projector

Sensor

IR LED IR LED

ToF depth camera Structured light depth camera

29

Page 30: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Distortion – Global Illumination

• Complex illumination affects the sensing performance (Global Illumination)

Sensor

Projector

Sensor

IR LED IR LED

ToF depth camera Structured light depth camera

30

Page 31: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Error – Specularity

• ToF depth camera

31

Page 32: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Error – Translucency

• ToF depth camera

32

Page 33: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Error – Specularity

• Structured light depth camera

33

Page 34: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Error – Translucency

• Structured light depth camera

34

Page 35: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Error Analysis

35

• Data collection & analysis

Page 36: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Error Analysis

ToF Sensor

36

Page 37: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Depth Error Analysis

Kinect Sensor

37

Page 38: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Color-Depth Calibration

Page 39: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Given a calibrated TOF-Stereo system

• Each TOF point PT defines a correspondence between PL and PR

Page 40: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Correspondences (samples) obtained by using the calibration parameters

• each correspondence comes from a TOF point

• different color -> different depth

Page 41: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Correspondences (samples) obtained by using the calibration parameters

• each correspondence comes from a TOF point

• different color -> different depth

Page 42: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

TOF-to-Left Mapping

• We use the left image as reference

Page 43: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

TOF-to-Left Mapping

Resolution mismatch

Page 44: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Left-to-Tof Occlusions

TOF-to-Left Mapping

Left-to-Tof Occlusions: the depth decreases from left to right

Page 45: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Tof-to-Left Occlusions

TOF-to-Left Mapping

Tof-to-Left Occlusions: the depth increases from left to right

Page 46: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Point Cloud filtering

• We reject points in left-to-tof occluded area

• We keep the minimum-depth points in case of overlap (due to Tof-to-left occlusions)

Page 47: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Disparity Map: Initialization

• Run Delauney-Triangulation on low-resolution point cloud

Page 48: 3D Imaging with ToF Camera - khu.ac.krcvlab.khu.ac.kr/CVLecture24.pdf · Tof-to-Left Occlusions: the depth increases from left to right . Point Cloud filtering •We reject points

Disparity Map: Initialization

• Run Delauney-Triangulation on low-resolution point cloud…

• …and initialize the stereo disparity map