Acoustic Sensor Network for Surveillance and Target Acquisition of Terrestrial Military Operations Gonc ¸alo Atan´ asio Academia Militar IST / UL Lisboa, Portugal Email: goncalo.atanasio@ist.utl.pt Ant´ onio Serralheiro CINAMIL, Academia Militar INESC-ID Lisboa, Portugal Email: antonio.serralheiro@academiamilitar.pt Ant´ onio Grilo IST / UL INESC-ID Lisboa, Portugal Email: antonio.grilo@inesc-id.pt Abstract—The main objective of this work is to present a system that detects and identifies combat vehicles with a wireless sensor network that contains acoustic sensors. The objective is to gather the sound from the combat vehicles, with acoustic sensors in the nodes of the wireless sensor network, perform the classification of the vehicle in the node and then send the packet to the base station using packet routing and data aggregation. The data base of vehicle sounds, drawn from combat vehicles of the Armed Forces, was made by training Gaussian mixture models (GMM) with data that explores the spectral characteristics of the sound, Mel Frequency Cepstral Coefficients (MFCC) and Beats per minute (BPM) were used as observation data. Vehicle detection and identification is performed with a GMM classifier. The NS3 simulator was used to simulate the wireless sensor network and the respective data aggregation and routing. The routing protocol used is DSDV, and it was tested over two different technologies, IEEE 802.15.4 and IEEE 802.11g. For the tests of the classifier, the number of Gaussian mixtures was optimized and the tests were made with different time windows for the MFCC. It was concluded that the classifier is better without the rhythm detection, that is, using only the MFCC it can obtain an F-score of 0.875 with eight combat vehicles on the data base. The simulations done using the NS3 simulator indicate that there must be a trade-off between the delays, the radio range and power consumption. The 802.15.4 technology, with aggregation, would be the most appropriate choice to the case study. However if it is desired to cover big areas, because of tactical restrictions, the most suitable technology is 802.11g. 1. Introduction With the evolution of the technology, wireless sensor networks have expanded their potential application domains. In many applications, a very high data traffic must be provided despite the limited hardware resources that leads to an unsustainable power consumption. Thus, low power consumption routing algorithms are needed as well as data aggregation, for the purpose of increasing the nodes lifetime. These nodes allow the acquisition, processing and send- ing/receiving data. When inserted in a military context, wireless sensor networks requires certain concerns such as resistance to changes in the network, redundancy, scalability, low power consumption and low latency for the identification and clas- sification of vehicles as close to real time as possible. Since the technology is available to both conflict sides, the advantage will go to those who best understand its scope and limitations. Today, the ability of detecting and classifying vehicles in an operational context have great tactic importance, since, increasingly, there is a technolog- ical evolution from the adversary and therefore, a higher speed of information flow is required for decision-making so, this technology allows you to optimize the command and control. Thus, with the proper application of this technology it is possible to increase the protection of military forces on the battlefield. Even with the existence of acoustic sensors in the Portuguese Army [1], the development of this tech- nology becomes important because the current systems do not allow detection and identification of military vehicles in an autonomous way which is a current limitation. In this particular case it will be applied a wireless sensor network to a military environment, whose purpose is the detection and classification of military vehicles in opera- tional context. These sensor networks usually consist in a set of stationary nodes distributed in a particular area. The implementation objectives are: Vehicle Sound Collection; Vehicle classification by processing the data in the node; Routing the result of classification to the base station; Using data aggregation in intermediate nodes in order to reduce network energy consumption. The system to be developed is summarily shown in Figure 1. In this figure is represented a vehicle approaching the wireless sensor network, which itself emits a certain sound that is collected by the closer nodes of the network. When a node collects a sound fragment through a micro- phone it performs its pre-processing by extracting spectral features. Each node has a database containing GMM models representing various vehicles, which is used to perform classification of the collected sound fragment. After the classification the node sends a packet with 1