Features and Signatures
In this chapter, we will discuss a important aspect in the core analytic of human re-identification, that is, converting a raw image into identifiable and reliable features.
- PDF / 930,557 Bytes
- 9 Pages / 439.37 x 666.142 pts Page_size
- 89 Downloads / 204 Views
Features and Signatures
In this chapter, we will discuss a important aspect in the core analytic of human re-identification, that is, converting a raw image into identifiable and reliable features.
2.1 Feature Detection Extracting representative features from objects is a useful way to detect and track targets. A feature is considered to be representative if it has expressive texture in a particular region. Also, a desirable feature should be invariant to changes in illumination and viewpoint. Harris and Stephens [1] proposed a corner detector known as the Harris corner detector. Let I(x, y) be the intensity of pixel (x, y) in a gray scale image I and s(u, v) be the weighted sum of squared differences obtained by a shift (u, v) of the patch. We have w(x, y) (I(x + u, y + v) − I(x, y))2 (2.1) s(u, v) = (x,y)
in which w(x, y) =
1 (x, y) is in the shifting patch 0 otherwise.
(2.2)
Expanding I(x + u, y + v) in a Taylor series around (u, v) = 0, we have I(x + u, y + v) ≈ I(x, y) +
∂I(x, y) ∂I(x, y) u+ v. ∂x ∂y
(2.3)
Then Eq. (2.1) becomes
© Springer International Publishing Switzerland 2016 Z. Wu, Human Re-Identification, Multimedia Systems and Applications, DOI 10.1007/978-3-319-40991-7_2
13
14
2 Features and Signatures
∂I(x, y) ∂I(x, y) u+ v − I(x, y) s(u, v) ≈ w(x, y) I(x, y) + ∂x ∂y (x,y) = xy H xy
2 (2.4) (2.5)
in which H is the Harris matrix denoted by 2
⎤ ⎡
∂I(x,y) ∂I(x,y) ∂I(x,y) w(x, y) w(x, y) (x,y) (x,y) ∂x ∂x ∂y ⎦ 2 H = ⎣
∂I(x,y) ∂I(x,y)
∂I(x,y) w(x, y) w(x, y) (x,y) (x,y) ∂x ∂y ∂y
(2.6)
where the partial derivatives can be estimated by applying Gaussian derivative filter to the image. A response function for the corner is given by R = det(H) − k · trace(H)2
(2.7)
in which k is a tunable sensitivity factor (usually between 0.04 and 0.15). The detector is more sensitive when k is smaller. R is large positive value if both of the two eigenvalues are large. A Harris corner is found if the response of a patch is locally maximum and higher than a preset threshold. Noble [2] introduced an alternative response measurement function amounting to a harmonic mean of eigenvalues from H as Rn =
2det(H) trace(H) + ε
(2.8)
in which ε is an arbitrary very small positive constant in order to avoid dividing by zero. Shi and Tomasi and Kanade [3, 4] improved this corner detector, known as the Kanade–Lucas–Tomasi (KLT) corner detector, so that the corners are more suitable for tracking. The KLT corner detector uses the criterion min(λ1 , λ2 ) > T
(2.9)
in which (λ1 , λ2 ) are the two eigenvalues of H, and T is a threshold. This way the corner detected by the KLT corner detector will be more reliable to be tracked over time. Harris and KLT corner detectors are effective, but may be computationally intensive due to numerous multi-scale/Gaussian convolutions or eigenvalue computations, which make them unsuitable for real-time high-resolution video sequence processing. Rosten and Drummond [5] proposed an algorithm called Features from Accelerated Segment Test, known as the FA
Data Loading...
