视觉里程计原理(一)特征提取(SURF算法)
- 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
- 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
- 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。
MPIG Seminar0045
Feature Extraction
陈伟杰
Machine Perception and Interaction Group (MPIG)
cwj@
Feature Extraction
Refined based on the book:
Mastering OpenCV with Practical Computer Vision
Projects_full.pdf
and
Bay H, Tuytelaars T, Van Gool L. Surf: Speeded up robust features [M]. Computer vision–ECCV 2006. Springer. 2006: 404-417.
or F for [R|t]Drawing path
The main steps of Visual Odometry
images parameters
Feature Extraction
Feature
matching
Compute E
First
Feature Extraction What feature is?
Characteristics can be easily identified in images
Edges Corners Blobs
lines points
Harris SIFT SURF Commonly used algorithm:
•Corner extractor
•Fast operation •Poor resolution •Not applicable when scale
changes •Blobs extractor
•Slow operation
•Good resolution
•Scale invariance
•Upgrade from
SIFT
•Speed up
•More robust
SURF(Speed Up Robust Feature) opencv2/nonfree/features2d.hpp
SurfFeatureDetectordetector()
SurfDescriptorExtractor
SURF(Speed Up Robust Feature)
Integral image
ii x,y = i=0i≤x j=0
j≤y
I(i,j)
(x,y)
A
C
B D
123
4
ii 1= A ii 2= A + B
D =ii 1+ii 4−ii 2−ii 3
SURF(Speed Up Robust Feature)
Hessian matrix
H x,σ=L xx(x,σ)L xy(x,σ) L xy(x,σ)L yy(x,σ)
L x,σ=Gσ∗I(x,y)Gσ=ð2g(σ)
ðx2g(σ) is Gaussian function andσis variance
It’s the image conversion like frequency domain transform
Approximation of Hessian matrix
Det
H
approx
=D xx D yy−0.9D xy2
L yy L xy
D yy D
xy
We can use integral image to compute easily now
H x,σ=
L xx(x,σ)L xy(x,σ)
L xy(x,σ)L yy(x,σ) Filter template
SURF(Speed Up Robust Feature)
scale space (image pyramid)
SIFT SURF Change size of filter only Easier and faster
Positioning feature points
3×3 window
“x”is extreme point when it’s the
max or the min of 26 points around
Setting a threshold value t
if x>t, x is feature point
The larger t is, the less points will be
SURF feature descriptor
Main direction
Statistics harr wavelet feature around the feature point with the range of 60°in a circle of radius6s(s is the scale of the feature point)
The max value is main direction
SURF feature descriptor
main direction
feature point 20s Every descriptor has
4*4*4=64 dimensional vector