Acoustic ENSBox: A System of Self-Calibrating Distributed Acoustic Arrays

Acoustic ENSBox is an ad-hoc deployable wireless system designed to support distributed acoustic sensing applications.  The platform is a self-contained unit containing an ARM-based CPU module, a wireless network interface, a 4-channel acoustic sampling interface, and a battery.  The system connects to a “head unit” that hosts an array of 4 microphones and 4 piezo tweeters in a Lucite and aluminum chassis.  The microphones are condenser microphones with a custom pre-amplifier board.  The photo at right shows an Acoustic ENSBox node deployed in the James Reserve.  The Acoustic ENSBox is not intended for long-term unattended deployments.  The system will run for about 24 hours continuously on a single 12V 7.2AH gel cell.  Longer term deployments could be achieved by duty-cycling the system.

In addition to the hardware, the Acoustic ENSBox includes a complete stack of system software designed to support distributed acoustic sensing.  The system autonomously forms an ad-hoc wireless network that supports inter-node coordination, hosts routing services and reports diagnostics to a user with a laptop.  It supports accurate time synchronized sampling, enabling application programmers to trivially compare time series data taken at the same time at two or more nodes.  An acoustic localization system (described below) autonomously and accurately estimates relative position and orientation for all nodes in the system.

With this stack of system software, this platform is ideal for many types of collaborative sensing, especially target localization algorithms based on “beam-crossing”, where multiple states estimate bearing to a target and combine their estimates to compute a location.  We hope to see the Acoustic ENSBox platform taken up by several groups at UCLA who are involved in acoustic localization projects.  We are currently working with Prof. Kung Yao's group to compare their bearing estimate algorithms to those developed for the position estimation application. We are also working with a student from Prof. Charles Taylor's group who is developing software on the Acoustic ENSBox platform to detect acorn woodpecker calls.

Wideband Acoustic Localization System

We performed early work in acoustic ranging using wideband audible acoustics (2000) and further developed this work as part of the GALORE project (2001-2002) and the SHM system at Sensoria (2001-2003).  In 2005 we started developing a new acoustic ranging system designed for the Acoustic ENSBox.  This capability is a critical part of its usefulness as a distributed acoustic sensing system, because the beam-crossing localization applications we want to develop require precise estimation of the 3-D location and orientation of the sensor arrays.

This work has been highly successful, resulting in a highly accurate, self--configuring localization system.  This system estimates the 3-D position and orientation of a collection of nodes, with no prior knowledge or anchor points.  The resulting relative coordinate system is then fit to anchor points if absolute coordinates are required.  The system works outdoors and, unlike many competing systems, is highly resilient to environmental noise and obstructing foliage.  In tests localizing 10 nodes in a forested, hilly region 70m x 50m, the system achieved an average of 20 cm position error.

Our system also leverages the 8cm baseline microphone array supported by the Acoustic ENSBox to estimate 3--D direction of arrival (DOA).  In controlled tests, we achieved an error distribution with a standard deviation of 0.96 degrees, and a maximum range of +/- 2 degrees. These results represent an improvement upon similar published work. 

The images below show some of the conditions of our James Reserve test.  The left image shows the approximate deployment laydown plotted on an aerial photograph of the James Reserve location (a UC Reserve located in Idyllwild, CA and managed by UC Riverside).  The right side image shows the conditions present between nodes 108 and 104.  The red box contains a magnification of the corresponding box in the image. As you can see the environment has some dense foliage and line of sight was partially obstructed.

The next two graphs show some of the results from our tests in the James Reserve.  The first two graphs show the accuracy of our Azimuth estimator under controlled conditions.  This component estimates the incoming direction of a ranging signal.  We see that the standard deviation of the estimates is 0.96 degrees while the total range of errors is 4 degrees.  We suspect that these errors may result from the fact that we do not calibrate the array geometry (i.e. each array’s microphones will be positioned slightly differently.

The next two graphs show the observed error in tests of zenith angle estimation (degrees above or below the horizon).  Negative angles perform poorly because the signal arrives from below the array and is significantly blocked by the base of the array.

The next two graphs show the performance of the ranging system.  We see that on the whole the ranges are within 5 cm and are not significantly affected as a function of distance.  Note that this data was scaled to compensate for temperature, but only a single temperature value was used even though the test was conducted in a semi-outdoor environment over the course of several hours.  We expect that some of the error is due to environmental changes.  Note that when we run the multilateration system we avoid this issue by taking all the measurements in a brief span of time during which we can assume that environmental changes were small.  In the distribution graph we show two curves.  The narrower distribution is formed by dropping all “outliers”: points with more than 10 cm of error.

Finally, the next graph shows the error observed in X/Y position.  For this graph, we ran 6 trials and computed position estimates for each trial. Then, we compared the X and Y position estimates for each node over the 6 trials.  In the graph below, each node is represented by a pair of X/Y error bars.  The error bars represent the distribution of X and Y coordinate estimates, considered separately for each node.  Then, to represent them all on a single graph, we subtracted out the “ground truth” positions, so that we only see the error or bias in the results.

Thus, if this system worked perfectly, all of the error bars would be very small and centered at (0,0).  Because the system is not perfect, the nodes are offset from the origin, and that offset represents the average bias in the position for that node.  From the graph, we can see that except for one node, they are all within about 10 cm of their ground truth locations (in fact, the average is 7 cm).  If we also include the Z axis, the average position error is 20 cm.  This is because Z is less well constrained, since the nodes are mostly in a plane.

More detailed information can be found in the PhD thesis.pdf  and in the powerpoint presentation .ppt .  Further publications are forthcoming.

Data

The data collected to test this system is available from our data repository at this link.

Software

All software used to build this system is located in the CENS CVS repository.

There is also some helpful information on using the system located in the CSL wiki.

·        Drivers for the VXP440 sound card are in emstar/sensors/audiod/vxpc_server

·        Time synchronization support is in emstar/timesync

·        The ranging system software is in emstar/devel/loc/ar

·        The multilateration software is in emstar/devel/loc/multilat

·        The StateSync multihop routing and coordination primitive is in emstar/devel/state

·        The diagnostic reporting and control software is in emstar/emproxy (emrun, emview, emproxy and rbsh)

Publications

L. Girod, M. Lukac, V. Trifa, and D. Estrin, "The Design and Implementation of a Self-calibrating Acoustic Sensing Platform". In Proceedings of the ACM Conference on Embedded Networked Sensor Systems (SenSys 2006), Boulder, CO. November, 2006. .pdf

L. Girod, "A Self-Calibrating System of Distributed Acoustic Arrays", Ph.D. Thesis, UCLA, 2005. .pdf  (Note that the results of the test in the forest have been improved since the thesis was published.  We found that the largest source of error in our results were inaccuracies in our initial survey of ground truth. After re--surveying, we achieved the results quoted above.)

L. Girod, M. Lukac, A. Parker, T. Stathopoulos, J. Tseng, H. Wang, D. Estrin, R. Guy, and E. Kohler, “A Reliable Multicast Mechanism for Sensor Network Applications”. Center for Embedded Networked Sensing Technical Report #48, April 25, 2005. .pdf

Hanbiao Wang, Lewis Girod, Nithya Ramanathan, Deborah Estrin and Kung Yao, "A Platform for Collaborate Acoustic Signal Processing" CENS Technical Report 00XX, November 28, 2004. pdf ps

W. Merrill, L. Girod, J. Elson, K. Sohrabi, F. Newberg, W. Kaiser, "Autonomous Position Location in Distributed, Embedded, Wireless Systems", In Proceedings of the IEEE CAS Workshop on Wireless Communications and Networking, Pasadena, CA, Sept. 2002. .pdf

Lewis Girod, Vladimir Bychkovskiy, Jeremy Elson, and Deborah Estrin, "Locating tiny sensors in time and space: A case study", .ps, .pdf In Proceedings of the International Conference on Computer Design (ICCD 2002), Freiburg, Germany. September 16-18 2002. Invited paper.

Jeremy Elson, Lewis Girod, and Deborah Estrin, "Short Paper: A Wireless Time-Synchronized COTS Sensor Platform, Part I: System Architecture" .ps, .pdf In Proceedings of the IEEE CAS Workshop on Wireless Communications and Networking, Pasadena, California. September 5-6 2002.

L. Girod, D. Estrin, "Robust Range Estimation Using Acoustic and Multimodal Sensing" (.ps) (.pdf) IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2001) , Maui, Hawaii, October 2001.

L. Girod, "Development and Characterization of an Acoustic Rangefinder" (.ps) Technical Report USC-CS-00-728, April 2000.

Powerpoint


girod@lecs.cs.ucla.edu
Last updated 20 May 2006