VoxNet: An Interactive, Rapidly-Deployable Acoustic Monitoring Platform Michael Allen Lewis Girod Daniel T. Blumstein Ryan Newton Deborah Estrin Cogent Computing ARC Samuel Madden CENS Coventry University MIT/CSAIL UCLA Abstract potentially target any of these applications, evaluated based on a specific bio-acoustic application. Distributed acoustic sensing underlies an increasingly The specific application we consider relates to the prob- important class of sensor network applications, from habi- lem of automated census of in-situ animal populations us- tat monitoring and bioacoustic census to security applica- ing bio-acoustic signals such as animal calls [6, 27, 9, 16]. tions and virtual fences. VoxNet is a complete hardware This application requires sophisticated localization, signal and software platform for distributed acoustic monitoring enhancement, and classification algorithms to process the applications that focuses on three key goals: (1) rapid de- acoustic time-series data, as well as statistical algorithms ployment in realistic environments; (2) a high level pro- to process the resulting animal call event traces. In some gramming language that abstracts the user from platform cases automated census may require combining acoustic and network details and compiles into a high performance detections with other sensor inputs such as imagers trig- distributed application; and (3) an interactive usage model gered based on timely results from the acoustic data pro- based on run-time installable programs, with the ability to cessing pipeline. While algorithms have been developed run the same high level program seamlessly over live or for census and population measurement for many specific stored data. The VoxNet hardware is self-contained and species [17, 6, 27, 9, 16], for some species and environ- weather-resistant, and supports a four-channel microphone ments this is quite challenging, and general solutions re- array with automated time synchronization, localization, main elusive. and network coordination. Using VoxNet, an investigator Because acoustic monitoring applications present many can visualize phenomena in real-time, develop and tune on- challenges that are not readily addressed by existing plat- line analysis, and record raw data for off-line analysis and forms [20, 19], previous work in acoustic applications has archival. This paper describes both the hardware and soft- typically required significant investment in platform devel- ware elements of the platform, as well as the architecture opment [28, 17, 5]. Several key challenges complicate the required to support distributed programs running over a development and deployment of acoustic applications: heterogeneous network. We characterize the performance • of the platform, using both microbenchmarks that evaluate Acoustic data is generated and processed at high rates, specific aspects of the platform and a real application run- placing a heavy computational burden on the sensor ning in the field. platform. Even with large memories and fast 32-bit processors, the processing pipeline must be highly op- timized to perform well. 1. Introduction • Developing on-line processing algorithms generally requires pilot studies to collect sample data. These pi- Acoustic sensing is a key component in a broad range lot studies are often similar to the eventual deployment of sensor network applications, including gunshot localiza- in terms of the logistical efforts and hardware function- tion [23], weapon classification [28], acoustic intrusion de- ality required. tection [5], biological acoustic studies [17, 30, 3], person tracking [25], speaker localization [8], and smart confer- • While some acoustic applications require on-line pro- ence rooms [29]. Existing work has described specific al- cessing, others require retrieval of complete traces, or gorithms, and in some cases custom hardware. This paper some balance between the two, perhaps dynamically is about a flexible hardware and software solution that can configurable at run-time. Both on-line processing and storage contribute to system load and impact platform 3. To define an interactive usage model based on run-time requirements. installable programs, with the ability to run the same high level program seamlessly over live or stored data. The design of our platform, called VoxNet, stems from considering these observations in the context of bio- 2. Related work acoustics applications, with a particular focus on animal Our related work is concerned with both platforms tracking and census problems, but with the potential to be for network tasking, and processing frameworks for high- applied in other domains thanks to a flexible, deployable frequency domains (acoustic, seismic). hardware platform and a high-performance distributable programming interface. 2.1. Platforms and architectures Software flexibility is required to support multiple con- Tenet [14] advocates a tiered networking approach to in- current applications, and in particular to allow reconfigu- network processing, where less capable platforms have ration and tuning of applications running in the field. For clearly defined roles and are tasked by more powerful plat- example, concurrently with continuously running an ani- forms. The reasoning behind this type of architecture is that mal call localization application, a researcher might want the complex aspects of a computation should stay on more to archive raw data or test out a new detection algorithm powerful platforms, because it is less costly and error prone without disturbing the running system. To achieve this we than push the computation to less capable platforms. implemented an interface for VoxNet that allows new pro- Tenet’s goals differ from ours. In the reference imple- grams to be installed in a running system without inter- mentation, Tenet nodes are Mica2/TelosB class devices, and rupting existing applications. Distributed VoxNet applica- micro servers are Stargate class processors. In contrast, our tions are written as a single logical program, abstracting the lowest-tier sensor nodes must be much more capable than programmer from the details of the network and particu- Mica or TelosB motes to deal with high-frequency data, and lar hardware platforms. These programs can operate over a we do not enforce a boundary between processing tiers. In combination of live and static data, residing in a distributed fact, one of our aims is to make this boundary more trans- system of sensors and backend servers. This model enables parent, so it can feasibly adapt to changes in application and users to tune and further develop applications during pilot environment. deployments, and enables the system to be used as an in- teractive measurement tool while it is deployed. This is im- VanGo [15] is a system which is designed to capture high portant for many short-term scientific deployments, because frequency phenomena using devices which are constrained it allows a scientist to immediately explore newly observed not only by processing capability, but also network commu- phenomena. Installing program updates at run time is also nication bandwidth. A focus of VanGo is therefore to sup- useful in other contexts such as security monitoring. port data reduction after suitable application-specific char- Usually, greater run-time programmability incurs a cost acterization. in performance. To address this, VoxNet builds on prior VanGo forms a program as a linear chain of filters, de- work designing the WaveScript [10] language and compiler, signed to reduce data from its original form into events of which we extended to support the sensor interfaces and net- interest. These filters can be enabled or disabled, and there work functionality of VoxNet. VoxNet is the first embed- are a library of different filters, such as FIR, event detection ded target for the WaveScript compiler, and developing the and classification. Again, our work focuses on a different VoxNet backend motivated many new features and opti- class of devices than VanGo, and the programming model mizations. Implementing our animal localization applica- provided is, accordingly, much more general. Whereas tion using the WaveScript programming language and op- VanGo is limited to a linear chain of filters, VoxNet allows timizing compiler results in a 30% reduction in processor an arbitrary dataflow graph and operators are not limited to load and 12% in memory usage, compared with the hand- filter semantics. coded C implementation used in a previous version of the application (see Section 5). Mate/ASVM [24] provides a framework for application Thus, the main contributions of this work are: specific virtual machines (ASVMs), to allow developers to created customized runtimes which are application depen- 1. To develop a platform capable of rapid deployment in dent. Network reprogramming is a key focus of virtual ma- realistic environments for bioacoustic applications; chines such as Mate. In our target domains, as well, phe- nomena are often not well-characterized, and application re- 2. To provide a high level programming interface that ab- quirements may vary over the lifetime of the network. How- stracts the user from platform and network details and ever, we choose native code compilation for our programs compiles into a high performance distributed applica- in order to achieve the level of performance required for in- tion; and tensive signal processing. Because VoxNet has a relatively fast wireless network, uploading new binaries on demand is tractable. The Acoustic