Lecture 5 Flashcards
What are the two main ways pose tracking is typically done? What do these entail?
Could be done with pose estimation(Get camera pose at each frame) or motion tracking(model object’s motion). The typical model consists of a measurement model, a motion model, and a combining filter.
What is a Kalman filter what does it assume?
A kalman filter is used to estimate the state of an object being tracked, it has a state we want to estimate, a measurement model, a motion model, and error terms for each. The Kalman filter assumes a linear state update model, a linear measurement model, random errors with gaussian distributions, known covariance of the errors, and predicts the state and overall covariance, trying to minimise the overall covariance.
What are the kalman filter equation steps?
Take the previous state estimate and covariance, use them to predict the next state and covariance, predict the measurement, and then update this estimate based on the real measurement.
What does the kalman gain act to do?
minimise the overall covariance.
What are two main variations of a kalman filter?
The extended Kalman filter, which uses Jacobians to allow for non-linear state updating of the measurement by making a linear approximation.
The unscented Kalman filter, does not assume known covariance of the measurement and equation. It estimates them as it goes by sampling points around the estimate.
What is a particle filter?
A particle filter is used to predict a state given we have some kind of measurement. It does this by producing a lot of potential states using a distribution, each one is a particle, These particles will each receive a weight, based on how they relate to the measurement, this is a measurement of how likely the particle is to be the correct state. We can then update likely states and check again.
What is the informal algorithm for a particle filter?
- Initialise n particles, pi
- Make a measurement, m
- Compute weights, wi
for each particle, based on agreement with measurement/ - Randomly pick particles based on wi
- Update their states (and add noise)
- Goto 2
What does sampling involve in a particle filter? What occurs after this?
In this step we resample particles, with high weight ones being more likely than low weight ones(particles can also be duplicated).
After this we update the samples using the state update function and adding random noise.
When will a Kalman filter be better than a particle filter?
When the particle filter algorithm would have been clustered, as this would essentially make it behave like a kalman filter.
Does a particle filter lead to the correct state only?
no, there isn’t a single estimate of the state, each particle is an estimate, what we should do with them depends on the situation.
