egaebel / lgtm Goto Github PK

@egaebel Thank you so much for your code! It helps me understand and implement SpotFi very much. But I have optimized your code to run more efficiently.

I am trying to use your experimental data to verify the correctness of my code. In your code, the antenna distance is set to be 0.1. To my understanding, if the antenna distance is larger than half length of the magnetic wave, it will course ambiguities of AoA detection in a linear antenna array. Thus when I run your original code, a time period of the Pmusic is:

This figure shows three paths in terms of ToFs. Each path appears multiple peaks. This confuses me a lot. Is it due to the 0.1 antenna distance coursing an ambiguity? Or am I missing the correct parameter setting?

Moreover, from your outputs, "True number of computed paths: 6". But the figure only shows three paths. Could you please help me explain it? (I had tuned the tau value to be 0-2000ns. It shows the shape of the figure above periodically. )

In lines between 646-652 of a file csi-code/spotfi.m:

for jj = 1:size(Pmusic, 2)
    % AoA loop
    for ii = 1:size(Pmusic, 1)
        Pmusic(ii, jj) = 10 * log10(Pmusic(ii, jj));% / max(Pmusic(:, jj))); 
        Pmusic(ii, jj) = abs(Pmusic(ii, jj));
    end
end

I think abs should be removed. Because MUSIC spectrum will have negative value at non-candidate in decibels , and taking abs to the negative value will make a non-candidate to a candidate.

If I get wrong, please let me know about it.

Regards.

hi， your code have some issue. when you constuct the noise space , the eigenvectors not be sorted.

@egaebel I found the likelihood is negative by using your code. i don't know the reason. Besides, Why you use SVM to train the likelihood weights. Can you explain the two problem in detail? Thank you!

According to the authors of SpotFi we need to calibrate the phase shift over the antennas due to internal processing/conversion (see https://bitbucket.org/mkotaru/spotfimusicaoaestimation/src/master/README.txt as done in ArrayTrack). I can't find any location in the code where you are applying this offset. Did you just omit that step?

I have a WiFi device which has four antennae , the frequence is 5G, and every antenna receive 54 subcarriers.
I 'm not sure how to get a smoothed csi on it. Can you give me a hand or just a little hint? Thank you : )

Hi egaeble, it seems that you use your android smartphone as the tx for measuring the CSI. Is your phone working at the injection mode? If so, how to do that?

@egaebel
Hi,
I have read your code about lgtm project, thank you for your efforts ! It helped me a lot.
A lot of experiments have been carried out, unfortunately，when I use my data from Intel 5300 NIC, the precision of AOA estimtion is far from satisfactory.
I noticed the process of phase sanitization algorithm which use spotfi_algorithm_1 algorithm.
I have some ideas about that. I construct a model to quantify the time delay.
τ = T_error + T_tof + T_antenna
τ represent the total time delay.
T_error represent the time delay induced by PDD、STO and other phase error.
T_tof represent thre time consumed by signal propagation from transmitter to receiver.
T_antenna represent the time consumed by signal propagation between two antennas, which can be expressed as T_antenna = (n-1)d*sinθ / c. d is the distance between two adjacent antennas and c is the speed of light.
τ_fit represent the linear fit result.
The aim of linear fit is to remove the effect of random time delay induced by PDD and STO.
when τ_fit <= T_error + T_tof, the angle information is retained.
when τ_fit fall in between T_error + T_tof and T_error + T_tof + T_antenna, I guess that the angle information will be destroy.
and, when τ_fit fall >= T_error + T_tof + T_antenna, the angle information will alse be destroy.
I want to know that whether my guess is right or not ?
Will the process of liear fit will comeout such a situation ? I call it "over" liear fit.
This question has pusseled me for a long time. I will be appreciated of you if you could give me a hand.
thank you.

I'm doing the same experiment, but I found that the phase abstracted form CSI is not stable, Can you tell me that in which device that carries the 5300 NIC in your experiment?
Thank you very much! @egaebel

Hi @egaebel. Thank you so much for sharing your code!

I have got a question in spotfi.m.

In line 61 you write
packet_one_phase_matrix = unwrap(angle(csi), pi, 2);
thus the packet_one_phase_matrix is assigned to be the unwrapped csi values of the first package.

In line 84 you write
sanitized_csi = spotfi_algorithm_1(csi, sub_freq_delta, packet_one_phase_matrix);
in the parfor loop. Thus the packet_one_phase_matrix is passed into spotfi_algorithm_1, which means the same packet_one_phase_matrix (that generated from the first package) is used for all packages.

Maybe I miss something, but I don't think this should be the case. Could you please explain a little bit?

Hey guys, I just found out that if I choose different sets of values for tau in the function"aoa_of_music",I get different results. Should tau be determined with regard to the physical distance between transmitter and receiver?

@egaebel Thank you for your share of this very useful code. If it's convenient for you, i'll be much appreciated if you can send me your thesis, thank you very much.