Elsevier

Ad Hoc Networks

Volume 114, 1 April 2021, 102438
Ad Hoc Networks

Low data regimes in extreme climates: Foliage penetration personnel detection using a wireless network-based device-free sensing approach

https://doi.org/10.1016/j.adhoc.2021.102438Get rights and content

Abstract

As far as low-cost deployment is concerned, wireless network-based device-free sensing (DFS) is of great interest and has successfully demonstrated the feasibility in Foliage penetration (FOPEN) target recognition. The classification accuracy of this technology is known to dramatically decrease in extreme climates where the received signals tend to be severely attenuated; while deep learning approaches have boosted performance, they only perform effectively when trained with large amounts of labeled data. Consequently, it is still unknown how to ensure reasonable detection accuracy in extreme climates where sufficient samples are difficult to obtain. To address this concern, we adopt two special measures for performance enhancement in this paper. One measure is to employ higher-order spectral (HOS) analysis to transform the time-domain signals into the bispectrum image representations, so that the shift to an image classification task could provide the advantage of using the existing Convolutional Neural Network (CNN) models. More importantly, the immunity of the approach against the unwanted clutters in foliage environments can be improved. The other one is to present an end-to-end Deep Learning Data Augmentation and Classification (DLDAC) model comprised of a Deep Convolutional Generative Adversarial Network (for data augmentation) and a SqueezeNet CNN backbone (for target classification), which can improve the classifier performance by using the augmented data on-the-fly. Thus, the negative impacts of low data regimes in extreme climates can be considerably accommodated. To evaluate the effectiveness of the proposed approach, comprehensive experiments are conducted on a real FOPEN dataset collected by impulse-radio ultra-wideband (IR-UWB) transceivers under three severe weather conditions. The experimental results demonstrate that even when only 300 training samples are taken for each type of target under every weather condition, the average classification accuracy of the proposed approach is still better than 92% in terms of distinguishing between human and other targets.

Introduction

Foliage penetration (FOPEN), in particular personnel detection, is mission-critical to a variety of applications, ranging from military to surveillance. However, the detection and classification of targets obscured by foliage is still a challenging issue. Since the conventional manual-based approach is enormously expensive and labor-consuming, many advanced techniques have been developed in the last few decades. In general, there have been two classical methods: radar system [1], [2], [3] and hybrid-sensor system [4], [5], [6]. Perhaps radar is the most well-known approach because it could accurately distinguish targets with a very low false alarm rate. However, the implementation and maintenance of such a system are incredibly high-cost. Although the alternative approach employing hybrid sensors can dramatically reduce the cost, the classification performance requires to be further enhanced. Due to ever-increasing requests for large-scale monitoring of forested regions at local and national levels, a low-cost approach with high accuracy and robustness is urgently needed for personnel detection in foliage environments.

As far as cost-effective sensing is concerned, a promising technology, known as wireless network-based device-free sensing (DFS), has attracted extensive attention [7], [8], [9]. This technology has been proposed as a way to fully utilize radio frequency (RF) transceiver for not only data communication but also target sensing, mainly because the presence of each target in the wireless network environment would cause the RF signals to have different patterns. By employing machine-learning technologies to identify and interpret these patterns, it is possible to recognize different types of targets. Consequently, it only demands an RF transceiver without any extra equipment, which indicates that it has the potential to reduce the total cost substantially. In recent years, many efforts have been made to explore DFS technology in indoor environments, such as target localization, detection and recognition. However, deploying this technology amongst foliage has not yet been fully explored. Currently, only a few attempts have been made to use impulse-radio ultra-wideband (IR-UWB) [10] transceivers for personnel detection in foliage environments [11], [12], [13], [14], [15], [16]. Since using DFS technology in the FOPEN domain is still in its infancy, there are still some critical issues to be tackled. In particular, one of the greatest challenges in this application is how to overcome data scarcity in extreme climates. Unfortunately, no attempt has been made to address this issue in the literature.

In a foliage environment, the variations of climate-induced effects, such as temperature, humidity and foliage density, could have a considerable effect on the propagation of RF signals in an unpredictable way. If the model is only trained on the dataset collected under normal weather conditions, it would be quite hard to achieve reasonable accuracy. As a result, it is desirable to collect labeled samples including all weather conditions. However, obtaining such a dataset in extreme climates remains a great challenge. Naturally, people have to operate in a foliage environment exposed not only to harsh weather conditions but also to great danger, which makes the collection work extremely hard. Moreover, since the occurrence probability of severe weather is low, the datasets are still limited even we could collect data samples under such conditions. Although the feasibility of the DFS-based approach for FOPEN has been evaluated under different weather conditions, its performance is heavily dependent on a large volume of labeled training data [15], [16]. If the training dataset is in low data regimes, machine learning algorithms usually get trapped into overfitting, which means that the performance is excellent for the training samples but relatively poor for the testing samples [17]. Consequently, classification accuracy can be dramatically degraded in extreme climates due to the limited availability of training samples. Therefore, how to guarantee reliable performance in extreme climates with an insufficient dataset is still an open question.

In this paper, impulse-radio ultrawideband (IR-UWB) signals are utilized for target recognition in foliage environments. Their bispectrum image representations are extracted as the discriminative features for not only making the input format more feasible to fed with the deep model but also minimizing the effect of performance caused by unwanted clutters. In addition, a Deep Learning Data Augmentation and Classification (DLDAC) model is presented to overcome the unavailability of sufficient labeled training data in extreme climates. The proposed DLDAC model is composed of an online data generation module and a classification module. Specifically, the Deep Convolutional Generative Adversarial Network (DCGAN) [18] that is popular in previous works for data augmentation is adopted as our data generation module. Also, the SqueezeNet [19] with less learnable parameters is employed as the classification module, so that the chance to overfit can be further minimized. Unlike several existing works [20], [21], [22] striving to implement data augmentation with GANs separate from classification, our DLDAC model is unique in several perspectives since it adopts an end-to-end learning scheme (i.e., jointly optimize the generation module and classification module): (1) The generated fake data in DLDAC can be fed into the classification module on-the-fly which does not need extra storage space to prepare generated data before training the classification module. (2) The diversity of these data can be largely improved since all the generated data are generated on-the-fly. Besides, the quality of the generated data can also be improved through information propagate from the classification module. (3) Our DLDAC is particularly designed for the task of FOPEN, which uses bispectrum images as input to deal with the limited data issue exposed in extreme weather conditions. Finally, an FOPEN dataset is obtained under three weather conditions, including rainy, snowy, and foggy, and these samples are then used to extensively verify the performance of the presented approach in terms of accuracy and robustness.

To sum up, the contribution of this work can be summarized in three-folds:

  • We present a DFS-based approach for FOPEN target recognition in extreme climates with training samples in low data regimes.

  • A combination of the bispectrum feature images and a well-designed DLDAC model is utilized to deal with the challenge of scarce training data in extreme climates.

  • Comprehensive evaluations are given to verify the performance of our approach. In particular, the robustness is further evaluated by variations of training samples and signal-to-noise ratio (SNR) values. Experiments demonstrate that the average classification accuracy of at least 92% is achieved on an FOPEN dataset collected under three severe weather conditions using a small number of training samples.

The rest of this paper is organized as follows. We first review the related works in Section 2. Section 3 introduces the experimental dataset taken in a foliage environment under several severe weather conditions. In Section 4, the details of the HOS-based feature extraction method are reported. The proposed DLDAC model for target classification is presented in Section 5. Experiment results and discussions are given in Section 6. Finally, the conclusion is shown in Section 7.

Section snippets

DFS-based FOPEN target recognition

As predicted, by 2030, there will be more than 100-billion devices connected through wireless sensor networks [23], [24], [25] to enable the Internet of Things (IoT) [26]. Since each of them must have a capability to transmit and receive RF signals, DFS perhaps would become the most promising sensing technology, which has the potential to dramatically reduce the overall system cost. Recently, followed by DFS-based approaches successfully applied for indoor target recognition, some researchers

Experimental dataset

To verify the feasibility of the presented approach, particularly the performance in extreme climates, measurements were conducted under three severe weather conditions, namely rainy, snowy, and foggy. The testing scene under such weather conditions is given in Fig. 1. All the samples were collected in a foliage environment dominated by a mixture of hardwood trees and low-lying undergrowth. Four target types are used for data collection, including (1) no target (without any placed target); (2)

HOS-based feature extraction

As far as FOPEN is concerned, due to the multipath propagation effect of RF signals, clutters generated by vegetation and rough ground are likely to overwhelm the scattering of targets that may include essential features. Consequently, it is more difficult to classify different targets as previously mentioned. As feature extraction is needed in this paper, higher-order spectral (HOS) analysis [35] is adopted to extract polyspectra in the collected RF signals. Since the polyspectra of a Gaussian

DLDAC for target classification

After transforming the original RF signals to bispectrum image representations, we propose the DLDAC model to improve the performance under the condition of limited training samples. The architecture of the proposed DLDAC model is shown in Fig. 6. There are two components involved in the deep architecture, i.e., the data generation and classification module. Taking advantage of an end-to-end learning scheme, the two modules are jointly optimized so that the small dataset can be enriched in a

Data split

In order to evaluate the effectiveness of our presented approach, we split the created FOPEN dataset into three subsets: training, validation, and testing. There are four different target types and three severe weather conditions involved in the dataset. As previously mentioned, under every weather condition, 1000 data samples are collected for each target type. To form the required training set for data-scarce condition, we only randomly select 300 samples for each target type under each

Conclusion

In this paper, the performance analysis of using the DFS approach for FOPEN personnel detection in extreme climates with insufficient training data is investigated for the first time. To ensure reliable performance in terms of classification accuracy, a DLDAC model is used in conjunction with HOS-based feature extraction. HOS analysis assists the raw signals to be converted into image inputs that are acceptable for the cutting-edge CNNs used for target classification. Furthermore, with the help

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Yi Zhong received the Ph.D. degree in information and communication engineering from the Beijing University of Posts and Telecommunications, Beijing, China, in 2017, and in computer science from the University of Technology Sydney, Sydney, NSW, Australia, in 2019. She is currently an Assistant Professor with the School of Information and Electronics, Beijing Institute of Technology, Beijing. Her research interests include signal processing, machine learning, and deep learning.

References (44)

  • LaouiraMohammed Lamine et al.

    An efficient WSN based solution for border surveillance

    IEEE Trans. Sustain. Comput.

    (2019)
  • ZhongYi et al.

    Multi-location human activity recognition via mimo-ofdm based wireless networks: an iot-inspired device-free sensing approach

    IEEE Internet Things J.

    (2020)
  • WangJie et al.

    Multi-target device-free wireless sensing based on multiplexing mechanisms

    IEEE Trans. Veh. Technol.

    (2020)
  • ZhangRonghui et al.

    Device-free wireless sensing for human detection: the deep learning perspective

    IEEE Internet Things J.

    (2020)
  • RenQingchun et al.

    Throughput and energy-efficiency-aware protocol for ultrawideband communication in wireless sensor networks: A cross-layer approach

    IEEE Trans. Mobile Comput.

    (2008)
  • ZhongYi et al.

    Device-free sensing for personnel detection in a foliage environment

    IEEE Geosci. Remote Sens. Lett.

    (2017)
  • ZhongYi et al.

    Internet of mission-critical things: human and animal classification—A device-free sensing approach

    IEEE Internet Things J.

    (2017)
  • ZhongYi et al.

    Impact of seasonal variations on foliage penetration experiment: A WSN-based device-free sensing approach

    IEEE Trans. Geosci. Remote Sens.

    (2018)
  • HuangYan et al.

    Cost-effective foliage penetration human detection under severe weather conditions based on auto-encoder/decoder neural network

    IEEE Internet Things J.

    (2018)
  • SrivastavaNitish et al.

    Dropout: A simple way to prevent neural networks from overfitting

    J. Mach. Learn. Res.

    (2014)
  • RadfordAlec et al.

    Unsupervised representation learning with deep convolutional generative adversarial networks

    (2015)
  • IandolaForrest N. et al.

    Squeezenet: Alexnet-level accuracy with 50x fewer parameters and¡ 0.5 MB model size

    (2016)
  • Cited by (0)

    Yi Zhong received the Ph.D. degree in information and communication engineering from the Beijing University of Posts and Telecommunications, Beijing, China, in 2017, and in computer science from the University of Technology Sydney, Sydney, NSW, Australia, in 2019. She is currently an Assistant Professor with the School of Information and Electronics, Beijing Institute of Technology, Beijing. Her research interests include signal processing, machine learning, and deep learning.

    Tianqi Bi received the B.S. degree in electrical engineering from the Beijing Institute of Technology, Beijing, China, in 2020. She is currently pursuing the M.S. degree in electronics and communication at Beijing Institute of Technology, Beijing, China. Her current research interests include signal processing, machine learning, and deep learning.

    Ju Wang received the Ph.D degree in signal and information processing from the Beijing Institute of Technology, Beijing, China, in 2004. She is currently an associate research fellow with the School of Information and Electronics, Beijing Institute of Technology, Beijing. Her current research interests include spread spectrum communication, satellite navigation and radar signal processing.

    Siliang Wu received the Ph.D. degree in electrical engineering from the Harbin Institute of Technology, in 1995. He then held a postdoctoral position with the Beijing Institute of Technology, where he is currently a Professor. His current research interests include statistical signal processing, sensor array and multichannel signal processing, adaptive signal processing and their applications in radar, aerospace TT&C, and satellite navigation. He has authored and coauthored more than 300 journal papers and holds 72 patents. He received the first-class prize of the National Award for Technological Invention and the Ho Leung Ho Lee Foundation Prize in 2014. He is also a recipient of the State Council Special Allowance, the National Model Teacher, the National May 1 Labor Medal, and the National Outstanding Scientific and Technological Personnel.

    Ting Jiang received the B.S., M.S. and Ph.D. degree in Communication and Information System from Yanshan University, PR China, in 1982, 1988 and 2003, respectively. Currently, he is working as a professor in Beijing University of Posts and Telecommunications. His research interests cover a wide range of topics in the wireless broadband interconnection, the information theory, the short distance wireless communication technological theory and application and the wireless sensor network.

    Yan Huang received the B.S. and the M.S. degree from the School of Computer Science and Engineering in Sichuan University, China, and Beihang University, China, respectively. He is currently a Ph.D. student with the Faculty of Engineering and Information Technology, University of Technology Sydney (UTS), NSW, Australia. His research interests include deep learning and machine learning.

    This document is the results of the research project funded in part by the National Natural Science Foundation of China (NSFC) under Grant 62071061 and in part by the Beijing Institute of Technology Research Fund Program for Young Scholars .

    1

    This is the first author footnote.

    View full text