Autonomous Guidance System for Weeding Robot

in Wet Rice Paddy

Choi, Keun Ha (14R55508)Department of Mechanical Engineering

Korea Advanced Institute of Science and Technology

Advisor In TITECH : Associate Professor Edwardo F. FUKUSHIMA/

FUKUSHIMA Robotics Lab.

Closing Ceremony 2014

’14. 8. 22

I. Introduction

Need to the robot in agricultureAging in agriculture population

Decrease the productivity and shortage of manpower

<2>

Weeding RobotsNeed to the robot, Problem of the hand weeder, riding weeder

Mechanization rate for weeding control 6.6%

Environment-friendly organic agricultural

[Hand weeder] [Riding weeder] [Weeding robot]

Autonomous

II. About Research

<3>

Autonomous Guidance System based-on IR Vision Sensor for Weeding Robot in Wet Rice Paddy- Robust to illumination condition- Improve the accuracy of rice row and guidance line + Low cost and Simple system

Process

Crop-soil(water)Discrimination

Crop row detection

Guidance line

Motion Control

Research progress

Crop row detection

[ SDI ]

[Single image / Static thresholding ]

Research results in KAIST

Research results in TITECH

II. About Research

<4>

System Architecture

NIRImage acquisition

Sequential difference Image (SDI)

Clustering rice pixels(K-means clustering)

Extract guidance lines

Gray images

Image subtract

Median filter

Thresholdering(OTSU method)

Crop segmentation

Evaluate guidance lines

Robot motion control

Yes

No

Extract the segmented crop line

Make the extended virtual line

Intersection distribution from lines

Line fitting(Robust regression) Robot guidance information

Research results in TITECH

Improving the algorithm

III. Concept of Algorithm

Concept : Rice Morphology Characteristic

<5>

Find the algorithm which is represented the convergence of central rice stem

• Step 1. Extract the segmented crop line from crop image

• Step 2. Extend each segmented crop lines Make the extended virtual lines

• Step 3. Compute the intersection from lines Intersection distribution

• Step 4. Extract the guidance lines from intersection distribution

• Step 5. Evaluate the guidance lines

Convergence line = Segmented crop line + Extended virtual line

Segmented crop line

Extended virtual line

Intersection distribution

Central stem

Convergence of rice stem : leaf originate from a central stem symmetrically

Step 1. Extract the segmented crop line from crop row image

<6>

Preprocess of crop row image

SDI image(binary image)

Remove small blob noiseSkeletonization

(Zhang-Suen thining)

Clustering using K-means algorithm

1, ... ,

| ( , ) 1

where k = centroid locations of initial cluster

k = the number of local peaks on sum of white pixels reference to

n

k

X x p i j

C c c

y

| ( , ) ( , ), 1, ... , k

( ), i = 1, ... , ki n n i n j

i i

X x d x c d x c j

C c X

( )1

4

( , )

10

N

n i nn

pre cur

pre

D d x c

D DD

D

( , )

( , )

( , )

| all pixels of

P 0 , If Bolb noise elimination

P 1, otherwise

, where number of pixels in the region

which is connec

area aream n k

areai j

areai j

P p P

P

P

ted each ( , ).p i j

( , ) pixel value is 1 (white pixel).i jp

IV. Process of Algorithm

0 100 200 300 400 500 600100

150

200

250

300

350

400

450

Image coordinate (x)

Imag

e co

ordi

nate

(y)

0 100 200 300 400 500 6000

20

40

60

80

100

Image coordinate (x)S

um o

f wih

te p

ixel

s

<7>

Segmented crop line

r

-80 -60 -40 -20 0 20 40 60 80

-500

-400

-300

-200

-100

0

100

200

300

400

500

a) Original image, (b) Skeletonized image, (c) Segmented lines (start point: circle, end point: cross, length: green line), (d) Extended virtual lines (start point: rectangle, end point: cross, length: blue line)

• Hough transform(Paul Hough, 1962)

- The classical Hough transform was concerned with the identification of lines in the image

IV. Process of Algorithm

<8>

Step 2 : Extend each segmented crop lines

Extended virtual line

∶

∶

∶

∶

Segmentedcrop line

Extended virtual line

,

,,

,

, if 90 ,

/ cos( ) , else

where , = length of segmented line

he

r s

r s

ll

l l

l e l

, ( cos( ), sin( )) if y and 0 / 2 , ( cos( ), sin( )) if y and / 2<, ( cos( ), sin( )) if y and

( ) =

ts ce te ce e e cs ce

ts ce te ce e e cs ce

ts cs te cs e e cs ce

p p p p l l yp p p p l l yp p p p l l y

L m

0 / 2

, ( cos( ), sin( )) if y and / 2<, ( , 0) if y and / 2, ( , 0)

ts cs te cs e e cs ce

ts ce te ce h cs ce

ts cs te cs h

p p p p l l yp p p p l yp p p p l

if y and / 2cs cey

[Segmented line] [Extended virtual line]

IV. Process of Algorithm

<9>

Step 3 : Compute the intersection from lines

Step 4 : Extract the guidance lines

1

20

0

0 0

0

0

01

1

0 0

0

0

,

0

0

0 0

0

0

0

0

0

0

0

0

0

0

0

0 0

0

Intersection distribution Robust regression (James O., 1988)

• Least squares estimates for regression modelsare highly sensitive to (not robust against) outliers.

Pixels = Moneybox Pixel which pass through the

extended Line Put money(‘1’) in a moneybox

Intersection point = {money > 1}

0 100 200 300 400 500 600 7000

50

100

150

200

250

300

350

400

450

500

Cluster 1Cluster 2Cluster 3

ti i iy x 1

ˆ ˆ0 [( ) / ]N ti i ii

x w y x

2( ) for |r|

0 for |r| >

ir

iew

ˆ ˆ( ) /ti i i i ir y x

ˆˆ Med | | /.6745ti i iy x / max( )i n n

where

IV. Process of Algorithm

V. Experiments

Results

<10>

0 100 200 300 400 500 600 7000

50

100

150

200

250

300

350

400

450

500

Cluster 1Cluster 2Cluster 3

(red: LSE, green: robust regression)

[Intersection point]

[Guidance line]

VI. Campus life in TITECH

<11>

Advisor

Travel

Friends

Healing!