object tracking Flashcards
provide some examples of tracking
CCTV camera, object recognition, speed measurement
What is the mean shift algorithm
is a statistical method, where the interest is to find the model in the distribution where the density is the highest, where it looks at the local highest density
how is the density calculated
the density is the weighted sum of the pixels in the neighbourhoods, for this particular case the weights are symmetrical
explain briefly how the mean-shift work
initially: finding the mode
the idea consists of computing the density and in order to find the mode, its gradient should be equal to 0, hence, the only way to nill it is when the centroid and centre are identical.
shifting: hence the main idea is that the window shifts towards the centroid on every iteration until convergence
what are some use cases for the mean shift
can also be used as segmentation utils or
how can we use the mean-shift for moving cells
for instance cancerous cells and how to deal with cell division
the idea is to initiate points offline and then let the mean-shift track the objects on every frame, simply ser the video backwards
what type of problems do we deal with when it comes to mean shift
if we have problems that can be translated to density problem then the mean shift is ver adequate
what is the motions estimation
generally speaking, motion estimation defines the way objects move in a set of frame and we can distinguished two :
corner based
pixel based
explain pixel-based motion
the idea is that we make the hypothesis that between every frame there is a displacement between pixels while the global intensity is different such that :
given two images the pixels are the same but with a different displacement :
hence we need to approximate (u,v) where use the gradient then taylor expansion
the idea is that we want to minimize the displacement values
what is the taylor expansion
how to express an image and the neighbourhood, using the initial image and the derivative of the image
what is the aperture problem and its fix
this problem occurred during the motion estimation where we do not have enough equations to represent the (u,v) map (displacement) where the gradeient doesn’t necessarily describe the motion
Lucas-Kanade,
the idea is that the neighbour’s pixels behave similarly, hence we define a window and
explain the particle filter work
initially, we define the grid, where we start with a particle which can be described as a set of parameters:
we spread the particles randomly on the image, then on each frame, we try to compute the similarities between the original image and the randomly positioned particle, then we order them from the most suitable to the least suitable one
how to deal with changing object size in a particle filter
the idea is that we simply add the scale as a parameter
how to deal with two moving objects in a particle filter
the idea is that we track one object and the other object we define the angle as a parameter
