STUDY OF THROUGH WALL IMAGING FOR TARGET DETECTION

Abstract

Through wall imaging (TWI) is one of the most rapidly emerging technologies where it tries to 'see' through visually opaque material like different types of walls and detect and image various targets behind the wall. It is a challenge for current researchers to design TWI system as well as interpret its data. The detection of targets becomes more challenging when no apriori information of walls and targets is available. TWI scene may consist of various types of targets with different shapes and material properties (dielectric). ThusTWI system should have the capability to detect, locate, classify the objects and should be able to obtain size and shape of objects present in room whichwill be useful to the end user for interpretation. It is well known that radar suffers with strong clutter problems. The signal received from radar consists of desired response of target with other signals arising mainly from radar system parameters, wall reflections, environment, and multiple reflections. The undesired component in the received signal is referred as clutter. In spite of tremendous amount of work by various researchers, it is difficult to select a suitable clutter reduction technique by which target detection accuracy can be enhanced and false target detection reduced. In TWI, with several indoor objects of different material and shapes, robust detection and classification is an important area of concern to end users. An important step towards classification of targets is thresholding. Several algorithms have been proposed by researchers for computing optimum threshold level in order to discriminate between targets and background. Thresholding becomes difficult when target and background level posses substantially overlapping distribution. Thus optimumthresholding technique is required to be explored in TWI images. There are different levels of obtaining information in TWI system depending on the application. Basic level which is called as 0-D provides information about presence of

target only. To get the location of the target, the next level 1-D is used which is also called as A-scan or range profile. But it does not indicate how many targets are present in cross range. B-scan system provides this information along with their locations. B-scan is a collection of A-scans recorded along scanning line. The highest level of information is provided by C-scanning which adds a third dimension of height to that of a B-scan system. C-scan is obtained from ensemble of B-scan. Height information may allow discrimination between targets having different dimension with price paid for more scanning time. In TWI, one of the major thrust in near future can be developing more efficient imaging algorithm for getting more information like size and shape of the target from C-scanning system. Imaging methods in through wall have gained wide attention. Many imaging algorithms have been developed for TWI. The most commonly used techniques in TWI for image formation are back projection, beamforming, tomographic, m-k, Kirchhoff s etc. These techniques can be analyzed further to find suitable technique for TWI applications. The purpose of obtaining images is to extract essential information from images which should be used for recognition of targets. Many techniques have been developed to recognize targets based on feature extraction. Features can be obtained by electromagnetic analysis, time spectral analysis and statistical features. Pattern recognition technique can also be used to identify and classify targets. Very less work has been reported on pattern recognition based object classification in TWI. So, the present thesis is an effort to detect and classify stationary targets present behind wall as well as to image targets for shape recognition. Therefore the main objectives are: • Critical analysis of clutter reduction techniques for detection of metallic as well as low dielectric targets • A novel approach to detect and classify targets based on statistical based techniques • Study of prevalent imaging techniques for through wall target detection • Application of pattern analysis techniques for shape recognition The thesis is divided in seven chapters in which the first chapter provides a basic platform of research work by presenting a brief introduction, motivation, research gaps, problem formulations and details of experimental TWI setup. Experimental setup is ingeniously assembled which consists of a SFCW radar in UWB range (3.95 GHz to 5.85 GHz) and a 2-D scanning frame. SFCW based radar consists of Rohdc & Schwarz Vector Network Analyzer (VNA) ZVB8 and pyramidal horn antenna having bandwidth of 1.9 GHz. The scanner is made of wood which is used to scan the radar in horizontal as well as in vertical direction. SFCW radar system possesses several advantages over impulse type of radar systems. One of the main advantages of the SFCW system is that many sources of time varying measurement error including frequency dependent magnitude and phase variations of connectors, cables, directional couplers and antennas can be removed through calibration. Data are collected using the designed TWI system for plywood and brick wall for different targets like low dielectric and metal of different shapes. The target considered for low dielectric constant material is Teflon. Different regular shapes like square, rectangular and circle are considered. Data has been collected by varying distances between TWI system and wall, between wall and targets to see the effect of positioning of TWI system on detection of targets. Before data are used for further processing, pre-processing techniques such as external calibrations and velocity corrections are applied. To do velocity corrections, knowledge of wall thickness and dielectric constant is required for which wall characterization has been done. In Chapter 2, the existing techniques and methods for TWI system are briefly reviewed. The chapter addresses the development in techniques for improving detection, imaging and classification of targets. Study and critical analysis of various clutter reduction techniques and its implementation on obtained data are presented in Chapter 3. Some of the existing methods for clutter reduction in TWI rely on background subtraction, time gating and spatial filtering. But these techniques have drawback. In background subtraction, clutter remains present if the data is not collected at exactly the same antenna positions and in addition it is not possible in real scenarios to collect data without target. Time gating is successful for targets which are far away from wall but for targets near to wall the target response overlaps in time domain and cannot be separated in time. A spatial filter is used to notch zero spatial frequency which represents wall reflections. This filter may subtract low frequency components of target as well due to practical design constraint of filter. Radar detects high dielectric discontinuities as presence of high contrast causes the reflected signal to be very

strong. The problem of distinguishing the targets becomes more difficult if the target is having dielectric constant near to that of medium of propagation. Thus in TWI if the scene consists of targets having low dielectric constant then important aspect will be to detect the target with low false detection. Presence of clutter and attenuation of signal deepen the problem further. Thus there is a need to provide a solution which can improve low dielectric constant target detection. Therefore, this chapter explores the applicability of different clutter reduction techniques based on statistical signal processing techniques like Singular Value Decomposition (SVD), Principle Component Analysis (PCA), Factor Analysis (FA) and Independent Component Analysis (ICA). To critically analyze different clutter reduction techniques two different data sets were collected in which in first data set a single metal target is kept behind wall and in second data set two targets, Teflon and metal are placed behind the wall. From first data set results, it is observed that metal target peak becomes visible significantly after clutter reduction and false target detection is minimized. The performance of clutter reduction techniques is compared on the basis of signal to clutter ratio. In second data set, the important aspect which is considered is detection of low dielectric constant target in presence of metal target. The raw B-scan image detects only metal target where as Teflon target is not observed due to weak reflection. When the raw B-scan image is processed using clutter reduction techniques, it is observed that only ICA is able to detect Teflon whereas other techniques fail to detect Teflon. Chapter 4, focuses on statistical based thresholding approach for target detection as well as a probability distribution function based algorithm for target classification. An important step towards detection and good classification result is to extract target information from raw data by reducing clutter to maximum extent. The focus is to classify a metal and low dielectric targets behind the wall. Existences of false targets are observed after clutter reduction technique. After clutter reduction, thresholding is applied to segment the target from the background. Thresholding becomes difficult when low dielectric constant target is used as background gray levels and some of the target pixel intensities are same. Deciding optimum threshold is still a challenge in segmenting target from the background which is essential for correct classification of targets. Existing thresholding methods like maximum entropy, minimum cross entropy, cluster analysis technique, Otsu and statistics based (mean+standard deviation) are critically analyzed. The performance of algorithms is evaluated by

computing two performance measures, sensitivity and specificity. Sensitivity is proportion of pixels correctly identified by algorithms as target and specificity is proportion of pixels correctly identified as non target. The higher the values of these two measures (close to one) the more accurate is the algorithm. It is observed that when threshold value is small, false target detection increases. But detection sensitivity of low dielectric constant is still not adequate. To improve detection of low dielectric constant target and to reduce false alarm further, modified statistical based threshold technique is proposed. It meets the user specified performance requirement that is sensitivity and specificity in order to find optimum value of threshold. In this chapter after thresholding target discrimination is approached as a classification problem. The statistics of thresholded images is obtained and is use to find probability density functions. Thus statistical method that characterizes radar data can be used to get the information about targets. In past probability distribution of various clutters has been modeled as Weibull, Rayleigh and normal from target detection point of view. But use of probability distribution to classify targets has been given less attention. In this chapter, image statistics after thresholding is evaluated on the basis of their probability distribution function and based on this results target are classified. To obtain statistical analysis of target pixels, various distributions like Normal, Rayleigh, Cauchy and Weibull are applied and best fit is chosen to model it. Probability density function is then obtained from them on the basis of Chi-squared goodness of fit test. It is observed that Weibull distribution fits both targets (metal and Teflon) more accurately than other three distributions. Once the distribution is known, parameters are estimated using maximum likelihood estimator and then the groups of pixels are labeled accordingly. The scale and slope parameter of Weibull distribution is obtained for metal and Teflon targets. Validations of proposed technique is carried out for different set of data where the parameters value are observed and class of target either high dielectric constant material or low dielectric constant material is classified. The objective of chapter 5 is to explore some of the existing imaging algorithm like back projection, frequency wave number (co-k) and delay sum techniques. Normal Bscan image depicts low resolution features in image. Since the goal is not only to localize the target but also to improve cross range resolution, synthetic aperture technique is applied. To compare different imaging algorithm, image quality is

measured using metric such as entropy and ratio of standard deviation to mean. Images due to different imaging techniques are obtained on B-scan experimental data. After applying imaging algorithms, a more focused image is obtained compared to Raw B-scan image in which no focusing algorithm is applied. It is observed that the brick wall causes an overall drop in coherence compared to case when plywood wall is used. The objectives of chapter 6 are (i) C-scan imaging for target detection and (ii) feature extraction and shape recognition using neural network. Instead of using C-scan data for three dimensional image formations, two dimensional images is obtained. This reduces time complexity of imaging algorithm. After observing range profiles the location of target is determined. Either the peak magnitude at target location is taken or energy of pulse obtained by target reflection is taken as value at each grid point in two dimensional images. The obtained image is enhanced using image processing techniques like filtering, interpolation, and thresholding and edge detection. The basis of method of the target detection is thresholding. Though the target can be clearly detected after clutter reduction, the image obtained is still of poor quality. Thresholding helps in enhancement of image quality. For edge detection simple Sobel operator is used. This process detects outlines of an object and boundaries between objects and background of the image. From the resultant image the centre of target can be approximately obtained. Though the target image does not corresponds to actual shape of target, much vital information about the target can be inferred. To enable reliable recognition of target shape, the essential information or features must be extracted. Features for two different targets should differ as much as possible. To recognize target from its shape, features are extracted using different methods like moment invariant, Fourier descriptor and waveform based technique. The features obtained by all these techniques for single image are combined to form single feature vector. These feature sets are used to train the network to recognize various shapes of targets of obtained data. A simple feed forward network with one hidden layer and one output layer is used. The number of nodes in output layer depends upon different types of shapes used. Since three shapes are used, three nodes are used in output layer. To train neural network on more data, features from images which are synthetically generated for different shapes with different dimensions are used. It is observed that the network is able to recognize shape of targets satisfactorily. xiv Chapter 7 presents the summary of contributions made in the thesis and future scope of work.