Time Synchronization in Acoustic Localization for

Transcription

1 Time Synchronization in Acoustic Localization for Mobile Open-Source Network Deployment Scott Almquist Worcester Polytechnic Institute Major Qualifying Project 18 December 2009 Embedded Digital Systems Group 102 Approved for public release - distribution is unlimited This work is sponsored by the Department of the Air Force under Air Force contract FA C Opinions, interpretations, conclusions and recommendations are those of the author and are not necessarily endorsed by the United States Government.

2 Acoustic Localization Determining the location of a sound source by using an array of sensors The difference in arrival times is used to calculate an approximate position of the sound source Sound Source t 1 t 2 Microphones Time Differences of Two Microphones Result in a Hyperbola 12/18/2009 2

3 Setup FWG114pv2 image. Retrieved October 13, 2009, from Openmoko freerunner image. (2008). Retrieved October 13, 2009, from Dell inspiron image. Retrieved October 13, 2009, from 12/18/2009 3

4 Time Synchronization Time synchronization errors are the greatest source of positional error Mic 1 Mic meters minimum error 12/18/2009 4

5 Audio Data and Timestamps Hardware Level Ideal Software Level Timestamps Time Time Time 12/18/2009 5

6 Audio Data and Timestamps Hardware Level Real Software Level Timestamps Time Time Time 12/18/2009 6

7 Time Synchronization Test Setup If the phones are equidistant from the sound source, the time difference of arrival should be zero 12/18/2009 7

8 Using a UDP Broadcast Message to Synchronize Time Synchronization Between Two Phones Using UDP Broadcast Message Method of Synchronization Difference in Time Reported by Phone Two Minus Time Reported By Phone One in Milliseconds Time of Test in Seconds Reported by Phone One 12/18/2009 8

9 Outline Goal Recording Audio Data Precision i Timestamps Filtering Timestamps Synchronize the Phones Conclusions and Future Work 12/18/2009 9

10 Outline Goal Recording Audio Data Precision i Timestamps Filtering Timestamps Synchronize the Phones Conclusions and Future Work 12/18/

11 Goal Improve the Time Synchronization Between Two Phones Can be sub-divided into two categories: Precision in retrieving an individual phone s clock Synchronization between two phones 12/18/

12 Outline Goal Recording Audio Data Precision i Timestamps Filtering Timestamps Synchronize the Phones Conclusions and Future Work 12/18/

13 Recording All Audio Data In order to improve the results further, we need to know that all audio data is being recorded Initiation tests indicate that this is not an unreasonable assumption Over a seconds we would expect 24,212,664 samples. Phone One sent 24,215,552 samples (1.1% more than expected) Phone Two sent 24,217,600 samples (2.0% more than expected) Playing a sine wave and viewing graphs of the recording results in no obvious discontinuities or phase shifts. It is assumed that audio data being recorded is consistent and timestamps are imprecise 12/18/

14 Outline Goal Recording Audio Data Precision i Timestamps Filtering Timestamps Synchronize the Phones Conclusions and Future Work 12/18/

15 Precision in Retrieving Timestamps Audio is processed in buffers Buffers are a consistent size (1024 Samples) Sampling frequency is a constant (44,100 hertz) 1024 Samples 44,100 Samples Second Milliseconds One measure of the precision of an individual phone s timestamps is a consistent millisecond difference between successive timestamps 12/18/

16 Consistency of Timestamps Frequency Histogram of Time Differences of Timestamps for Phone One Method: gettimeofday Priority: Normal 60% within 1 millisecond Time Differences in Milliseconds Consistency in the ALMOND project was poor 12/18/

17 Consistency of Timestamps Frequen ncy Histogram of Time Differences of Audio Buffers for Phone One With Increased Priority Method: gettimeofday Priority: High 98.98% within 1 millisecond Time Difference in Milliseconds Increased priority was required for good results 12/18/

18 Advanced Linux Sound Architecture Audio library and kernel level API Default audio library for Openmoko Used in ALMOND project to record audio AL provides function calls to retrieve timestamps 12/18/

19 Consistency of Timestamps Frequenc cy Histogram of Time Differences of Timestamps Created Using AL API Function Calls for Phone One Time Difference in Milliseconds Results are bimodal, but it looks like they average to the correct value 12/18/

20 Outline Goal Recording Audio Data Precision i Timestamps Filtering Timestamps Synchronize the Phones Conclusions and Future Work 12/18/

21 Using a Moving Average Freque ency Histogram of AL Timestamp Differences Using a Moving Average of Two for Phone One Method: AL Priority: High 99.88% within 1 millisecond Timestamp Differences (in milliseconds) Using a moving average results in expected time differences 12/18/

22 Can We Lower the Priority? Frequenc cy Histogram of AL Timestamp Differences Using a Moving Average of Two and Regular Priority for Phone One Method: AL Priority: Normal 98.17% within 1 millisecond Timestamp Differences (in milliseconds) AL timestamps allow us to lower the priority and achieve similar results 12/18/

23 Moving Averages Freque ency Histogram of Time Differences Using AL Timestamps and a Moving Average of 10 Differences for Phone Two Method: AL Priority: Normal 99.78% within 1 millisecond Time Differences (in milliseconds) Moving averages improve the results, but outliers are still present 12/18/

24 Moving Averages Frequenc cy Histogram of Time Differences Using AL Timestamps and a Moving Average of 100 Differences for Phone Two Method: AL Priority: Normal 99.95% within 1 millisecond Time Differences (in milliseconds) Moving averages can be extended for better results 12/18/

25 Moving Medians Freq quency Histogram of Moving Median of Ten Time Differences Using AL Timestamps for Phone Two Method: AL Priority: Normal 99.95% within 1 millisecond Time Difference (in milliseconds) Moving medians have good results with fewer points 12/18/

26 Outline Goal Recording Audio Data Precision i Timestamps Filtering Timestamps Synchronize the Phones Conclusions and Future Work 12/18/

27 UDP Initial Synchronization uency Freq UDP Initial Synchronization Time Difference Values Difference in Time of Arrival at Phone Two From Phone One (in milliseconds) Average UDP Initial Synchronization is Milliseconds 12/18/

28 Outline Goal Recording Audio Data Precision i Timestamps Filtering Timestamps Synchronize the Phones Conclusions and Future Work 12/18/

29 Conclusions Consistency of Retrieving Time on an Individual Phone Method of Retrieving Timestamp Priority Moving Average Moving Median gettimeofday Normal None 60 gettimeofday High None AL High/Normal None <1% AL High Moving Average of AL Normal Moving Average of AL Normal Moving Average of AL Normal Moving Average of AL Normal Moving Average of Moving Median of 10 Percent of Time Differences within 1 Millisecond of Expected Value 12/18/

30 Conclusions UDP Synchronization can achieve an average synchronization of milliseconds between the phones. Future options for synchronization between phones include: UDP Synchronization GPS (Not with the Openmokos) NTP 12/18/

31 Future Work Prove that all audio data is being retrieved Synchronize time between phones A ti t? Are timestamps even necessary? 12/18/

32 Acknowledgements ALMOND Partners: Daniel Skehan Muhammad Saleem WPI Advisors: Professor Clancy Professor Heineman Lincoln Laboratory Advisors Albert Reuther Glenn Schrader 12/18/