如何实现视觉跟踪的粒子滤波器？

Question

So I just did sebsastian thrun's course on AI. In there, he mentions how to build a particle filter for tracking a moving xy robot based on heading theta and forward movement.

The code is here: https://gist.github.com/soulslicer/b4765ee8e01958374d3b

In his implementation, he does the following:

1. Get Range from Sensor of all bearings after moving R=1, Theta=0.5
2. Move all the particles by R=1, Theta=0.5
3. Compute the weights of all particles ranges against the measured range from sensor
4. Resample and draw new particles

---

This works for a motion model great. How on earth would this work for computer vision tracking? For example, I want to track a yellow circular blob. How would I "move" the particles? What might my cost functions be? Especially the moving part, I'm not sure how I'd do that step for computer vision tracking

---

Here is how I think it might work, but I'm probably wrong:

1. Get features from image, and compute the optical flow velocities of each feature
2. Place alot of particles in the scene with varying x,y,xvel,yvel
3. For the computation of weights, we can compare the each particle's velocity and position against all features
    If we can threshold out the object based on color/shape, can match image features to shapes and put that in the cost function
4. Resample and draw new particles

Answer 1

To use particle filtering, you need:

• a transition model (eg, a motion model used for moving the robot) and
• an observation model (ie, a model used to compute weights given sensor readings).

it is also helpful to clearly define spaces for

• observations (eg, a range for sensor readings)
• tracking states (eg, a range for robot positions)

Now, based on the description in your question, I'm assuming the goal is tracking positions of the yellow blob based on computed optimal flow features. Then I would model

• transition as a function that samples a new position given previous position by sampling noise only, eg, imagine using only + random.gauss(0.0, self.turn_noise) or + random.gauss(0.0, self.forward_noise) part in def move(self, turn, forward):
• observation as a function that returns high-scores for likely observation and state input pairs

The problem I see is defining the observation model, a likelihood function between blob positions and optimal flow outputs, is not trivial/intuitive, eg, is the yellow blob likely be in the center of high optical flow outputs area? if so how can I express such relationship as a likelihood function? For this reason, I would look into using different observations, eg, using outputs from noisy yellow blob detector.

My answer is based on pg 16. of the particle-filters.ppt file available at http://www.probabilistic-robotics.org/