Introduction

Challenges:

• Differences in cam viewpoints & orientations
• Light conditions
• Pose variability
• Change of Cloth

Related work

POV

• Match among images
• Biometric
• Color
• Some limited by movements of people(for people/car)
  • have method of maintenance [2]

Brief

• *** Very efficient description of object using local features, communication, point correlation
• Procedure:
  • target => idef & des by local features
  • des => wavelet(signature)
  • signature =Communicate=> neighboring camera node => gen des(specific signature)
  • Mathing algorithm

System

Architecture of proposed system:

Control component

• Input: query(set of images, from each camera)
• Reidef (Query image from op) => point detection & description

Reidef component

Interest Point Detection

Hessian matrix => minima maxima(using sign of Det) \[H(f(x,y)) = \begin{bmatrix} \frac{\partial^2 f}{\partial x^2} & \frac{\partial^2 f}{\partial x \partial y} \\ \frac{\partial^2 f}{\partial x \partial y} & \frac{\partial^2 f}{\partial y^2} \end{bmatrix}\]

Intensity Image, second derivative <= convolution with gaussian kernel(derivative,laplacian), scale \(\sigma\) \[H(x,\sigma) = \begin{bmatrix} L_{xx}(x,\sigma) & L_{xy}(x,\sigma) \\ L_{xy}(x,\sigma) & L_{yy}(y,\sigma) \end{bmatrix}\] (highlight ed)

Approx, weighted box filter => \(det=D_{xx}D_{yy}-(0.9D_{xy})^2\)

Accurate Interest Point Localisation

• Thresholding
• non-maximal suppression
• localisation => intrepolating & fitting (Brown [3])
  • Taylor expansion of H & d = 0
  • \(\hat{x} = -\frac{\partial^2H^{-1}}{\partial x^2}\frac{\partial H}{\partial x}\)

Feature Descriptors

64 & 128¹ dim vector with hue information, divide \(4\sigma *4\sigma\) \[[\sum d_x, \sum d_y, \sum |d_x|, \sum |d_y|]\]

The orientation is selected by rotating a circle segment covering angles of \(\pi / 3\)²

Point Correlation

Matching process, op give interest point \(x_j\), from frame interest point \(c_i\)

Sum of Quadratic differences(SQD)