codehime/mlsegment

Optical Flow

MLsegment

Abstract

Visual surveillance requires the moving object to be captured and tracked. The basic technique used
for discriminating moving objects from the static objects is background separation. Background
separation will subtract the current frame from the reference frame of static objects. This method is
not capable of handling multiple backgrounds like moving water in a fountain. Hence we have to
use some other models like unimodal model, MOG, Kernel , CB, etc. for separating those variations.

Team members

   1401001 - Deep Parekh
   1401011 - Jeet Parekh
   1401035 - Krima Doshi
   1401061 - Maharshi Bhavsar

Introduction

Foreground-background segmentation is required when motion of objects need to be detected. The moving objects can be considered as foreground and the stationary objects can be considered as background. Next, we used MOG filter for the initial separation. Then we took a reference frame and subtracted current frame from the reference frame. This approach works for the stable camera and a stable background and only a moving foreground. For the moving background-foreground segmentation we used a complex algorithm called Optical Flow. Optical Flow is one of the tool at rescue to solve many problems such as 3D shape acquisition, oculomotor control, perceptual organization, object recognition and scene understanding. Our problem statement is concerned with real time video sequence in which objects(Human) or the camera maybe moving according to 3D path. Optical flow provides visual perception.

Optical Flow

Optical flow is implemented by studying the velocity of objects. The velocity is related to the space-time image derivatives at one image location using an equation often called the gradient constraint equation. n. If one has access to only two frames, or cannot estimate It, it is straightforward to derive a closely related gradient constraint, in which It(x, t) in is replaced by I(x, t) I(x, t + 1) I(x, t).

The Algorithm

The Algorithm Optical flow can be implemented by following the below mentioned steps:

1. Compute the intensity of each pixel
2. For each pixel position compute the gradient matrix and store an eigenvalue of matrix
3. Separate the high scoring pixels by flag matrix F and region size k and flag region size f
4. Take the top n eigen values and use those for the trackable features
5. The Gaussian random distribution is applied for speed base
6. Next, warp one image, take derivatives of the other so you dont need to re-compute the gradient after each iteration
7. Repeat until complete