Elsevier

Signal Processing

Volume 92, Issue 3, March 2012, Pages 829-840
Signal Processing

Audio based solutions for detecting intruders in wild areas

https://doi.org/10.1016/j.sigpro.2011.10.001Get rights and content

Abstract

This paper presents an overview of the work that has been done in the field of wildlife intruder detection using only acoustic sensors. The motivation of such an application is related to protection of large wildlife regions, such as forests, lakes, and other natural reservations. The sounds of interest are represented by humans, engines, birds and animals. In order to simulate various environmental situations, different types of noisy environments have been considered. Both low complexity and standard audio classification methods are presented. Standard audio classification methods prove to be more robust, but at an expense of significantly increased complexity. Since low complexity systems are more feasible for monitoring remote areas, the complexity issue is discussed and solutions are proposed.

Introduction

The problem of sound classification has been addressed many times in the open literature and several solutions have been proposed in different fields. Medical applications, like classification of heart sounds [1], hearing aids [2] or remote monitoring systems [3] are very popular these days. Different solutions for environmental sound classification applications were proposed in [4]. Also, vehicle identification using wireless sensor networks [5] is a promising topic, with different applications in real life. Classification of combustion sounds is another interesting subject, with real life usability [6]. Probably the most known topic in which sounds classification has proved to be successful is the speech/speaker recognition field [7], [8], [9].

Humans play a critical role in ensuring the integrity of our forests and wild places. There are many natural reserves with wildlife, flora, fauna or features of geological or other special interest which are spread and there is practically impossible a continuous surveillance of all these areas. Although these regions are protected by law they are quite often the target of bad intentioned people for hunting, forest cutting and other. Moreover, the simple disturbance of wildlife by curious people could harm the endangered species. Not only the terrestrial reservations are the target of this illegal activities but also the protected lakes or coastal regions (such as the Danube Delta) are places of illegal fishing, or hunting of bird species strictly protected by international laws. At the same time, deforestation is proceeding at an unprecedented rate all over the world. The big problem is represented by the illegal deforestation which is hard to be detected in real time and stopped. Overall, systems of monitoring wild regions and detecting intruders are very necessary these days and would probably ensure a better preserving of these protected areas [10].

In the wildlife protection applications, standalone systems must be utilized. These systems will only send emergency messages to some remote observation station and must include all the classification and detection steps. This is because many areas under surveillance are remote and difficult to be visited and also the intrusion of people for surveillance should be limited. As one can see, the implementation of a wildlife surveillance system requires the design of low complexity algorithms and the utilization of hardware with low power consumption.

The work presented in this paper can be structured in two sections. In the first part we present a solution with low computation cost which could be easily implemented on a simple controller. In the second part, more complex approaches are utilized, increasing the accuracy of the system but at the same time increasing the complexity and the costs of the implementation.

This work is focused on performances of a potential monitoring system that could be assimilated to an acoustic eye. Its usage against other monitoring systems like the video surveillance ones has some advantages: simplicity of implementation, less information to be processed, it does not depend on the ambient lightning and represents a much cheaper solution. The ultimate goal of our work is to develop an acoustic sensor network which, placed inside a protected wildlife area, would be able to detect and classify several sounds of interest. The sounds of interest are related to several different events that must be monitored inside such protected areas. For the purpose of the work presented here we are interested in the detection and classification of only few sound classes: sounds originated from humans, birds, cars and animals [11].

Recently we have proposed different solutions for the problem of wildlife intruder detection [10], [11], [12], [13]. The studies in [10], [11] present a solution that uses TESPAR (Time Encoded Signal Processing and Recognition) as a method for sound encoding and classification. The work in [11] has improved the results from [10], especially when the S matrices have been used in the classification process. This proved to be very important; indeed, using only the S matrices, this would lead to a decrease in the complexity of the algorithm, which can be crucial in a standalone system with low power consumption. Even though there was also an improvement in the classification rates when various types of noisy environments were simulated, the rates were not fully satisfying. In [12], [13] we have proposed two standard sound classification methods, which proved to be more robust. Both of them were using as features the Mel-Frequency Cepstral Coefficients (MFCCs), while for training and classification were used Gaussian Mixture Models (GMM) and Support Vector Machines (SVMs). The motivation of trying complex approaches was threefold:

  • 1.

    We wanted to compare the results of these two different approaches, to see exactly how well does the low complexity method perform in comparison to standard sound classification methods;

  • 2.

    Taking into consideration the improved results of the standard classifiers, we suggest a combined solution, which utilizes both of these approaches;

  • 3.

    Finally, even though our purpose is to realize a system that has to be used in wildlife, our method can also be used for property surveillance. In this different situation low power consumption should not be mandatory anymore, thus more complex and robust algorithms can be utilized.

The rest of the paper is organized as follows. The theoretical backgrounds of this paper are briefly presented in Section 2. Technical design and implementation details are provided in Section 3. Section 4 presents the experimental results.

Section snippets

Theoretical background

In the following we shall briefly recalled some relevant properties of intrusion detection systems, TESPAR encoding, Mel-frequency cepstral coefficients, Gaussian mixture models and Support Vector Machines.

Design and implementation

In the following we shall briefly describe the databases used in implementation, then the two frameworks are presented: low-complexity experimental setup and standard classification experimental setup.

Experimental results with standard TESPAR

In this section we present the results for the initial tests [10], performed with TESPAR, and without the modification that was later proposed. When using S matrix we received an overall correct classification rate of 95.3%, while for A matrix the rate was 97.3% (Experiment 1.1). These results can be slightly improved (Experiment 1.2) by the band-pass filtering process (Table 1).

Besides the above mentioned base-line tests we have done several other experiments in order to verify the performance

Conclusions and further developments

In the work presented here we have evaluated the performances of a potential monitoring system that could be assimilated to an acoustic eye. In all the experiments, closed set identification was performed. Both complex and low-complexity solutions were employed and the results are compared. The low-complexity approach proved to be fairly robust when various noise environments were simulated. As one may expect, the standard sound classification methods presented proved to be more robust than the

Acknowledgment

This research was partially supported by CNCSIS project number 162/2008.

References (27)

  • S. Chu et al.

    Environmental sound recognition with time-frequency audio features

    IEEE Transactions on Audio, Speech and Language Processing

    (2009)
  • S.S. Yang, Y.G. Kim, H. Choi, Vehicle identification using wireless sensor networks, in: Proceedings of IEEE...
  • E. Lupu, M.V. Ghiurcau, V. Hodor, S. Emerich, Combustion sound classification employing Gaussian Mixture Models, in:...
  • Cited by (0)

    View full text