Re: [STDS-802-11-TGBF] 答复: [STDS-802-11-TGBF] 答��: Discussion for 20/12

Thread Links	Date Links
Thread Prev	Thread Next	Thread Index	Date Prev	Date Next	Date Index

Hi Rui Du,

Thank you for the clarification.

Regarding your response below, in monostatic scenario, no feedback from sensing receiver is needed since sensing Tx and Rx are in the same device. In bistatic scenario, it looks like you assumed the first tap corresponds to LOS, which can’t be always true in general. Without LOS information or a common reference (time or location) between a sensing transmitter and sensing receivers, different reflection environments may generate the same/similar TCIR (e.g., two taps only for a range of interest) for each sensing receiver or among sensing receivers at different locations. In this case, the device (sensing Tx or initiator) that receives and processes such a feedback can’t tell the difference. This is one of the consistency issues I mentioned before.

Best regards,

Ray

From: durui (D) <ray.du@xxxxxxxxxx>
Sent: Monday, January 24, 2022 9:02 PM
To: Rui Yang <Rui.Yang@xxxxxxxxxxxxxxxx>; STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx
Subject: 答复: [STDS-802-11-TGBF] 答复: [STDS-802-11-TGBF] 答复: Discussion for 20/1288 (TCIR) and the motion

Hi Rui,

Thanks for your email and please find my answers inline as follows.

Best wishes,

Rui Du

发件人: Rui Yang [mailto:Rui.Yang@xxxxxxxxxxxxxxxx]
发送时间: 2022年1月25日 4:49
收件人: durui (D) <ray.du@xxxxxxxxxx>
抄送: STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx
主题: RE: [STDS-802-11-TGBF] 答复: [STDS-802-11-TGBF] 答复: Discussion for 20/1288 (TCIR) and the motion

Hi Rui Du,

Thank you for the clarifications. I have a few follow up questions in the email below.

Best regards,

Ray

From: durui (D) <000017788cb650b9-dmarc-request@xxxxxxxxxxxxxxxxx>
Sent: Monday, January 24, 2022 3:04 AM
To: STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx
Subject: [STDS-802-11-TGBF] 答复: [STDS-802-11-TGBF] 答复: [STDS-802-11-TGBF] 答复: Discussion for 20/1288 (TCIR) and the motion

Hi Rui,

Thanks for your email and please find my answers inline as follows.

Best wishes,

Rui Du

发件人: Rui Yang [mailto:Rui.Yang@xxxxxxxxxxxxxxxx]
发送时间: 2022年1月22日 6:17
收件人: durui (D) <ray.du@xxxxxxxxxx>
抄送: STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx; Ron Porat <ron.porat@xxxxxxxxxxxx>
主题: RE: [STDS-802-11-TGBF] 答复: [STDS-802-11-TGBF] 答复: Discussion for 20/1288 (TCIR) and the motion

Dear Rui Du,

Could you explain how you defined “estimation accuracy”, shown in the figures you attached?

[Rui Du] It is true that the range resolution of 20 MHz is 7.5m in the simulation. But please note that the definition of resolution is different with accuracy (both of them are key indicators for a sensing systems).

The resolution is defined as the ability to separate two targets.

The accuracy is defined as how good a target can be estimated, it is usually defined as the RMSE/MSE between the estimated value and ground truth.

Btw, If you have a look at the 11bf use case document, you could find there are accuracies and separations (resolution) for each individual use case. And usually, the value of accuracy is (much) smaller than that of the separation(resolution).

[RY] My question is about your definition of “estimation accuracy”.

[Rui Du] I used the typical definition of ‘estimation accuracy’ describe above.

‘The accuracy is defined as how good a target can be estimated, it is usually defined as the RMSE/MSE between the estimated value and ground truth.’

In 20MHz channel bandwidth case, as you used in your simulation, the range resolution is 7.5 meters. If the “estimation accuracy” is defined as the difference between true location and estimated location of an object, how can you get such a high estimation accuracy (~0 meter in high SNR) without knowing the location of the object in advance? Thanks.

[Rui Du] The y axis should be labeled as estimation accuracy (range cell) rather than estimation accuracy (m).

[RY] Could you explain what is “range cell” and, for example, if estimation accuracy = 0.1 “range cell” (replacing “m” to “range cell” in your figure), what does it mean?

You may misunderstand my last question above. Let me try another way: in your PDP plots in slide 6 of 21/1288r4, what does “0” value in Range mean? In other word, what is the reference of “range”?

[Rui Du] Range cell is a typical sensing term and it is same with ‘range resolution’. As I described in last email, when estimation accuracy (cell) achieves 0, it means the target is detected at the range cell/range resolution where ground truth located. In the simulation, at each SNR level, 20000 time independent simulations are conducted, and an ‘estimated range’ and corresponding ‘range accuracy (cell) ’ can be got for each individual simulation. The ‘estimation accuracy (cell)’ in the simulation represents the mean value of accuracies from these 20000 simulations. 0.1 is the mean accuracy of 20000 times simulations, and it means the mean accuracy (as defined above) at this SNR level is 0.1 range cell/resolution.

Ok, now I get your question clearly, thank you for the rephrasing. Please find my answer as follows.

In the simulation, we are considering a simple case in which transmitter and receiver are fully synchronized, the ‘0’ in range is the time (could be transformed to range with tie ) when sensing PPDU is transmitted.

In real systems, this is could be solved with the information of LOS. Some commons methods are listed here, for example,

In monostatic sensing, the coupling between transmitting antenna and receiving antenna could generate a peak with very large value in measured CIR, this peak usually is regarded as ‘0’ in range.

In bistatic sensing, it is very hard to get range ‘0’ precisely with commercial WLAN system. And, the CIR will also be affected by some other effects (i.e. synchronization offset, etc.). Although the range ‘0’ is not available in this case, but as we discussed in previous email, the propagation difference between LOS and reflection path remain the same.

Thanks very much for the important question, and I will correct it in the following discussions/contributions.

In this simulation, we are only evaluating the sensing performance at range cell level. So, when accuracy(cell) reaches 0, it means that the target is detected correctly within the range cell where ground truth located. And of course, further signal processing could be applied to analysis the sensing accuracy(m) it could achieve within a resolution cell (limited by CRLB, and a similar performance gap between TCIR and grouped CSI still existing).

But, please note that, even with 20MHz, you are still able to achieve very high sensing accuracy (m) (~0 meter) at high SNR level when target is already separated, if some estimator (e.g. ML) is adopted. Many conclusions can be found from relevant academic papers.

[Rui Du] Btw, if you want to estimate a target, the first thing you need to do is create a target in your simulation. Definitely the parameters of the target (ground truth) are known in the simulation, because these parameters are adopted to simulate target. The signal is reflected by the target and received at the receiver and further processing is applied to estimate the target’s parameter.

This is a common way to evaluate the sensing performance , both for simulation and commercial systems.

For commercial system, the ground truth (e.g. information from GPS, etc.) is also needed during the test to evaluate the sensing performance.

Then, the system will be applied in actual scenarios. And of course in the practical scenarios, you may not have the actual ground truth. However, the system detection performance has already been validated!

In the simulation, we can achieve very good performance at high SNR level. In actual systems, it is usually very hard to achieve this kind of performance because of non-ideal effects (e.g. clutter, impairment of systems, RCS fluctuation, etc.). Simulations can take into account these non-ideal effects to a certain extend. But it might be difficult to completely model the real world non-ideal effects in simulations!

But please note although the sensing performance will reduce in real system, the performance gap between TCIR and grouped CSI still existing ! This is clearly proved by the theoretical bound (CRLB).

Best regards,

Ray

From: Ron Porat <000009a0da80e877-dmarc-request@xxxxxxxxxxxxxxxxx>
Sent: Friday, January 21, 2022 2:25 PM
To: STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx
Subject: Re: [STDS-802-11-TGBF] 答复: [STDS-802-11-TGBF] 答复: Discussion for 20/1288 (TCIR) and the motion

Hi Rui,

Thanks very much for the quick answers. At this point I don’t recall the exact discussions in 11ac MU-MIMO feedback.

I generally think that an SNR range>10dB is more applicable to us in reality and negative SNR as shown in the plots below are of no interest. Same for 20MHz BW, to the extent that it makes a difference it’s better to see results for 160MHz BW or at least 80MHz, at the least you will get more feedback tones even with Ng=16 for 160MHz BW.

In addition with frequency domain feedback the sensing application can calculate time domain if it so desires.

Also, I’m not following why range estimation accuracy is the metric you chose for a sensing scheme that looks at channel variation.

Also, if feedback overhead is a concern then sending an NDP is very efficient.

Thanks,

Ron

From: durui (D) <000017788cb650b9-dmarc-request@xxxxxxxxxxxxxxxxx>
Sent: Friday, January 21, 2022 3:49 AM
To: STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx
Subject: [STDS-802-11-TGBF] 答复: [STDS-802-11-TGBF] 答复: Discussion for 20/1288 (TCIR) and the motion

Hi Ron,

Thanks for your email and please find my answers inline as follows.

We can have an online discussion if needed.

Best wishes,

Rui Du

发件人: Ron Porat [mailto:ron.porat@xxxxxxxxxxxx]
发送时间: 2022年1月21日 2:32
收件人: durui (D) <ray.du@xxxxxxxxxx>; STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx
主题: RE: [STDS-802-11-TGBF] 答复: Discussion for 20/1288 (TCIR) and the motion

Hi Rui Du,

Can you compare TCIR to decimated frequency domain with the same total number of complex samples fed back?

[Rui Du] Attached please find the simulation results for the sensing performance of two different targets (a strong target and a weak target).

In the simulation, the legend ‘Full CSI’ represent TCIR and has the same number of complex samples with the legend ‘grouping factor equals to 16’(which represent Ng-16 is adopted during the grouping).

Please note that the simulation is just an example and the feedback size of TCIR can be further reduced according to the applications. For example, only 1 tap (the tap with largest magnitude) is used in the experimental validation for the finger tracking (the results can be find in DCN 0660/1288).

In the simulation the x-axis is SNR and y-axis is the range estimation accuracy (defined as the root mean square error between the estimated range and ground truth). Both results (sensing for strong target and weak target) indicate that the estimation accuracy based on TICR (legend with ‘full CSI’) is better than the sensing performance based on same numbers frequency CSI (legend with ‘grouping factor equals to 16’).

Also, why would the Rx work harder to generate TCIR if frequency domain is good enough? FYI already in 11ac (e.g 1131) time domain was discussed as channel feedback but was not accepted.

[Rui Du] This is a very good question and I think we need to discuss how can we define ‘good enough’.

First of all, the simulations have shown that the sensing performance of TCIR is better than that of grouped CSI. This also can be explained by the CRLB I mentioned in previous email. Experimental validation of TCIR for finger tracking also has been presented in my previous contributions. So, based on the above mentioned, I personally think that TCIR could achieve better sensing performance than grouped CSI.

Maybe another relevant question here is ‘if frequency domain CSI is good enough already’ ?

To be honest, I am not sure at this stage. In WLAN sensing, the most important information we want is the feature(s) of intended targets in a given environment. In different applications, the target could be a human being , a pet or even a hand (i.e. hand gesture recognition). So the signal reflected from the target could be very small and also much smaller than LOS signal (this is quite different with communication which mainly relies on the LOS signal in most cases). So, if we want to sense a unknown target (which may have very low SNR), definitely it is good to adopt a method that has ‘better performance’ rather than a method has ‘good performance’.

Also, many WLAN sensing academic papers has been published based on the CSI estimated at the receiver. To achieve better performance, the researchers always would like to use all the subcarriers rather than part of the subcarriers.

Thanks for your information. But I think the advantages of TCIR in sensing is different with that in communication. If possible, could you please share me more information about why it was not accepted in communication ?

Regards,

Ron

From: durui (D) <000017788cb650b9-dmarc-request@xxxxxxxxxxxxxxxxx>
Sent: Thursday, January 20, 2022 12:35 AM
To: STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx
Subject: [STDS-802-11-TGBF] 答复: Discussion for 20/1288 (TCIR) and the motion

Hi Rui,

Sure, please take your time and also please find my answers inline as follows.

Best wishes,

Rui Du

发件人: Rui Yang [mailto:Rui.Yang@xxxxxxxxxxxxxxxx]
发送时间: 2022年1月19日 7:44
收件人: durui (D) <ray.du@xxxxxxxxxx>
抄送: STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx
主题: RE: Discussion for 20/1288 (TCIR) and the motion

Dear Rui Du,

Thank you for your long responding email. If you don’t mind, I will try to address your comments/questions in a few days.

Meantime, may I ask if the result with “Full CSI” in Figure 2 corresponds to the full PDP or full CIR without truncation?

[Rui Du] The legend ‘Full CSI’ in the simulation in DCN 1288 represents truncated CIR (TCIR). In the simulation, two targets are located at 30m and 120m, respectively. The feedback size of TCIR , corresponding to the range of interest .

Please note that the parameter adopted in the simulation is just an example. According to the 11bf use case document, the maximum range for WLAN sensing is 20 meters (i.e. store sensing). This means the feedback size of TCIR can be further reduced.

Do you have the results with different truncation length (different CIR feedback size)? Thanks.

[Rui Du] If the TCIR of can be selected/truncated properly from the ‘full CIR’, the increase of TCIR feedback size will not affect the sensing performance.

This is mainly because the sidelobe is affected by the number of subcarriers adopted during the transformation from frequency domain CSI to time domain CIR. Since all the subcarriers are used by the IFFT to generate the ‘full CIR’, the ‘full CIR’ has low sidelobe levels, which will improve the sensing performance(as the simulations indicate). The only problem is that how the ‘full CIR’ can be truncated properly.

To select/truncate the TCIR properly, two criteria shall be met(as we described in the answer for Q3 in last email).

[1] The index or position of the TCIR shall be selected properly. This can be solved by a good reference tap(e.g. the tap with largest amplitude).

[2] The size of the TCIR shall be designed properly. As figure 3 in last email shows, the delay difference Δt between the largest magnitude tap (reference tap) and the following tap represent the propagation delay difference between LOS path and reflected path. So, in different applications, the size of the TICR is decided by the range of interest in each application (the calculation method is also provided in my previous contribution).

For example, the range of interest can be set to 5 meters(or a little bit longer) if the information within 5 meters is need. The most important thing is that the subset corresponding to 5 meters should be selected properly(as the criteria mentioned above), and any information beyond 5 meters is redundant in this case. If you want more information, you could adjust the range of interest according to your application.

In summary, if the index/position of the TCIR is selected properly and the feedback size is enough to cover the range of interest, the increase of TCIR feedback size will not affect the sensing performance.

Please let me know if you have any further questions or comments.

Best regards,

Ray

From: durui (D) <000017788cb650b9-dmarc-request@xxxxxxxxxxxxxxxxx>
Sent: Sunday, January 16, 2022 9:12 PM
To: STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx
Subject: [STDS-802-11-TGBF] 答复: Discussion for 20/1288 (TCIR) and the motion

Hi Rui,

Thanks for your comments and please find my answers inline as follows.

Best wishes,

Rui Du

发件人: Rui Yang [mailto:Rui.Yang@xxxxxxxxxxxxxxxx]
发送时间: 2022年1月13日 22:39
收件人: STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx
主题: Re: [STDS-802-11-TGBF] Discussion for 20/1288 (TCIR) and the motion

Hi Rui Du,

Thank you for reaching out for comments. I apologize for missing your comment collection before.

Regarding the motion (#49) you proposed, I have a few concerns:

It is not clear to me why the frequency domain CSI is not sufficient for sensing? If the concern is about feedback overhead, it would be better to consider additional reduction based on the existing compressed CSI feedback. To generate the time domain CSI, the sensing receiver may need additional processing power, which would not be desirable.

[Rui Du] I am not very sure if the ‘compressed CSI feedback’ you mentioned is the compressed beamforming matrix(the precoding matrix) in the main stream of 802.11 or not. But if you mean compressed beamforming matrix here, the conclusion from TGbf is that it cannot be adopted for WLAN sensing directly, and that is the main reason why TGbf want to reuse the CSI matrix as a feedback type. If you have some ideas about how can we use compressed beamforming matrix in WLAN sensing, I think the group will be very happy to discuss it.

So, in this email, I am only focus on the evaluation of frequency domain CSI (CSI matrix) and CIR for sensing, and explain the reasons why we need the CIR.

It is true that the feedback overhead of frequency domain CSI can be reduced with greater grouping factor Ng, but (T)CIR could provide more advantages than overhead reduction! Here is one example figure (assuming 5 subcarriers) describes the approach of frequency domain CSI grouping.

Figure 1

[1] It should be noted that full subcarriers is better than grouped subcarriers for sensing theoretically. This can be easily understood if we think about the CRLB (Cramér-Rao Lower Bound), which is the theoretical bound describes the lower bound of MSE in the parameter estimation. The CRLB of delay can be calculated with the following equation.

For simplicity, we are discussing a single antenna case and the in the above equation is the SNR level at the receiving antenna. is the root mean square bandwidth of the signal used for sensing and can be calculated with equation . It is easy for us to know that the of full subcarriers is greater than of grouped subcarriers, which means the CRLB of full subcarrier is lower than that of the grouped subcarriers. That is the theoretical reason why full subcarriers is better than grouped subcarrier in sensing.

[2] Based on the feedback procedure of frequency domain CSI, the full subcarrier CSI is estimated and obtained at sensing receiver. Then the full subcarrier CSI will be grouped and fed back to sensing transmitter, which means the sensing transmitter can perform sensing only with the grouped subcarriers CSI.

In the simulation of DCN 21/1288, the (T)CIR we presented is generated with full subcarrier rather than grouped subcarrier. The result indicates that the sidelobe generated with full subcarrier is lower than the sidelobe generated with grouped subcarrier. The sidelobe level is very important in sensing for further detection and parameter estimation. One simulation result from 21/1288 is pasted here as figure 2.

Figure 2

In figure 2, the x-axis is the SNR level and y-axis is the estimation accuracy. The simulation results show that the estimation accuracy improves with the SNR increases. More important is that, the results clearly indicate that the sensing performance decreases with grouping number Ng increases. The simulation results are consistent with the theoretical analysis we mentioned above.

[3] It is the feedback overhead reduction. The feedback overhead of TICR is much smaller than the Ng can be used for now(Ng can be further increase for overhead reduction but the performance will further reduced, as the analysis listed above).

Based on the advantages described above on, TCIR could achieve better performance (and lower overhead) if the target is located in the feedback subset of the CIR. The problem that how the subset of CIR should be selected is explained in the answer for Question 3.

The first sub-bullet of the motion text (“Calculating the CIR (time domain) from frequency domain CSI through IDFT(usually, IFFT)” ) appears an implementation. Is your intention to mandate this operation at sensing receivers?

[Rui Du] I am not preventing any other methods or techniques that can be used to generate CIR. If you have any potential suggestion, we could discuss it.

Based on my understanding, FFT/IFFT is definitely a good choice based on two reasons.

[1] FFT/IFFT is widely adopted in the wireless communication already due to its low complexity and some other advantages, it is a good choice that we can reuse it.

[2] FFT/IFFT is an linear integral transformation. The adoption of FFT/IFFT usually won’t add any distortion to the signal and remain the information that is needed for further sensing processing. For example, angle information can be estimated based on the multiple CIRs (generated from frequency domain CSI by IFFT) from multiple antennas.

In the second sub-bullet of the motion, you have “complex samples corresponding to the range of interest of the entire CIR”. I am not sure how the “entire CIR” is defined and if it can be defined consistently among different sensing receivers. In addition, from your early contributions, it clears to me that this set of complex numbers (i.e., PDP?) may vary a lot depending on the algorithm at the sensing receiver (e.g., grouping size and timing relationship among those numbers). So, the question is how you can guarantee the consistency of the measurements from different sensing receivers.

[Rui Du] In the contribution, the entire CIR is defined as the output of IFFT of the full subcarrier CSI. And as we stated in the contribution, CIR may be affected by the synchronization offset. An example is shown in figure 3, solid line is the amplitude of the 1^st CIR and dash line is the amplitude of the 2^nd CIR. The amplitude of the 1^st CIR and 2^nd CIR may be different because the time synchronization is performed per PPDU. The important thing is that the relative delay Δt between the 1st peak and 2nd peak remain constant in different CIRs in a sensing device within coherent time. Physically, Δt represents the propagation difference between LOS and reflection path (R1+R2) as figure 4 shows. So, with some ideas borrowed from 60GHz, the consistency of the measurements ((T)CIRs) from single device can be achieved by selecting the subset with a reference point (e.g. the tap with largest amplitude). That means, no matter how CIRs is affected by the synchronization offset, the subset can be chosen around the tap with largest magnitude, and the consistency can be achieved.

I am a little bit confused with the ‘consistency of measurements from different receivers’ ? Can you explain why we need the consistency from different receivers ? So I can do further clarifications.

If you mean ‘how can we fuse the results from different sensing receivers ?’ The answer is that this is another complicated problem which is not related to what I proposed. But, please bear in mind that, If we have multiple devices in the sensing, the consistency of measurements per device can be achieved by the method we mentioned above. The CIRs measured at multiple receivers are different due to the different positions of receivers (e.g. assuming on transmitter and multiple receiver in figure 5), because they are coming from different sensing links. The important thing is to make the measurement per device consistent.

Figure 3

Figure

Figure 4

Figure 5

In general, I believe we should not ask sensing receivers, which need to feedback sensing measurement results, to process CSI for a measurement that we cannot ensure the consistency of its meaning and reference value over time and among different sensing devices.

[Rui Du] Based on my understanding, it is easy for the measurements to achieve consistency for each individual device with some selection rules mentioned above. And what I proposed here is trying to maximally utilize the signal for sensing with a very simple processing.

[Rui Du] Please let me know if you have any further comments or questions.

Best regards,

Ray

From: durui (D) <000017788cb650b9-dmarc-request@xxxxxxxxxxxxxxxxx>
Sent: Tuesday, January 11, 2022 10:44 PM
To: STDS-802-11-TGBF@xxxxxxxxxxxxxxxxx
Subject: [STDS-802-11-TGBF] Discussion for 20/1288 (TCIR) and the motion

Dear all,

I am sending this email to initiate a discussion on the contribution 21/1288 Truncated Power Delay Profile(Truncated Channel Impulse Response) - follow up and motion 49 I ran yesterday.

Actually, I already tried to collect as much opinions as possible from the group member before the motion, and I thought I’ve discussed thoroughly with the group members who feedback their concerns or questions. I was surprised with the motion results.

Hence, I would like to discuss if there is any further concerns. Please feel free to let me know (by 11bf reflector or private email) your thoughts and any suggestions for the contribution and the motion.

Best wishes,

Rui Du

To unsubscribe from the STDS-802-11-TGBF list, click the following link: https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-11-TGBF&A=1

Re: [STDS-802-11-TGBF] 答复: [STDS-802-11-TGBF] 答���: Discussion for 20/1288 (TCIR) and the motion

Re: [STDS-802-11-TGBF] 答复: [STDS-802-11-TGBF] 答��: Discussion for 20/1288 (TCIR) and the motion