New paper on #ArXiv!
"Tracking moving objects through scattering media via speckle correlations"
With @YJaureguiS, and Harry Penketh.
Short(?) thread explaining what it is about.
1/
arxiv.org/abs/2202.10804
"Tracking moving objects through scattering media via speckle correlations"
With @YJaureguiS, and Harry Penketh.
Short(?) thread explaining what it is about.
1/
arxiv.org/abs/2202.10804
Scattering is a major problem for imaging, and it doesn't take very much of it before we can't see essentially anything of what it is happening.
Not surprisingly, imaging in the presence of scattering is a very active field of research.
2/
Not surprisingly, imaging in the presence of scattering is a very active field of research.
2/
https://twitter.com/j_bertolotti/status/1481215310178967554?s=20&t=Y7yXVQG90DQF4NuY-7FPqg
There isn't a single best way on how to deal with scattering, and the answer depends a LOT on how much scattering we are talking about and its properties.
As a rule of thumb, the most complicated situation is where all light is multiply scattered.
3/
As a rule of thumb, the most complicated situation is where all light is multiply scattered.
3/
That said, linear scattering can't destroy the information, it "just" scrambles it. The result is that the scattered light forms complex patterns, which might look random but contain a lot of correlations where the information is hidden.
Which gives us some hope.
4/
Which gives us some hope.
4/
An approach that has attracted some attention is based on a correlation that links the angle of incidence of the incoming wave, with the output angle of the scattered light, called the "optical memory effect".
5/
5/
https://twitter.com/j_bertolotti/status/1285225731782520832?s=20&t=Y7yXVQG90DQF4NuY-7FPqg
To make the long story short, this correlation allows you to measure the convolution between the (unknown) object to be imaged, and the (unknown) speckle pattern generated by the scattering medium.
6/
6/
This by itself is not very useful (too many unknowns) but, as realized by Labeyrie in the '70s, performing an autocorrelation one can separate the information about the object from the information about the speckle.
7/
7/
The only remaining problem is that we are left with the autocorrelation of the object, not an image of the object itself.
This autocorrelation can be numerically inverted to obtain the desired image, but this is a slow and computationally intensive process.
8/
This autocorrelation can be numerically inverted to obtain the desired image, but this is a slow and computationally intensive process.
8/
What if things are moving? In principle one can perform an autocorrelation inversion per each frame, but this needs to be done in post-processing (too slow to do otherwise), making real-time tracking complicated.
9/
9/
What if we are more interested in the tracking itself than in the real-time imaging? (i.e. we care more about knowing how the stuff in the scene moved than the precise look of the scene)
In that case we show that there is a simple and computationally inexpensive solution!
10/
In that case we show that there is a simple and computationally inexpensive solution!
10/
Instead of autocorrelating the light collected at each frame, we cross-correlate the light detected at different times, and perform the difference between two such correlations (to subtract all the extraneous signal)
[Showing here a simulation. The experiment is in the paper]
11/
[Showing here a simulation. The experiment is in the paper]
11/
This results in a easy-to-compute video where the moving object is fixed in the middle, and the background moves around it, showing clearly what it is happening.
[Technical note: This allows for tracking well beyond the memory effect range.]
The end 🙂
12/12
[Technical note: This allows for tracking well beyond the memory effect range.]
The end 🙂
12/12
Paper accepted! 🥳
• • •
Missing some Tweet in this thread? You can try to
force a refresh
