Research Topic

Knowledge Based Target Detection for Through the Wall Imaging

Typical Radar B-Scan Image

Introduction

Through the Wall Radar Imaging (TWRI) is an evolving technology that gained much  attention in the last decade. It allows to sense through visually opaque building material and man made structures using electromagnetic wave propagation which has  numerous applications in civil engineering, search and rescue, cultural heritage diagnostics, law enforcement, and military applications. TWRI is faced with many  challenges, including detection and classification of a large variety of possible indoor targets   in presence of multipaths and unwanted wall signal attenuation and dispersive effects.  In addition, when considering stationary targets, Doppler signatures and change detection  techniques become ineffective and one has to perform detection and classification in the  image domain, as a post-processing step to beamforming.  The radar images of stationary targets obtained through a wall are subject to strong artifacts, which could visually appear as targets in intensity and spatial concentration. Human interactive systems won’t work in such environment. To avoid false alarms, robust computer-based  systems should be sought out and applied.

Image Formation

After having obtained the measurements, the first step in through-the-wall radar imaging is the image formation or beamforming. In this step the measured raw data is processed such that a 2D cut (also called B-scan) or a 3D image of the scene is generated.

High resolution radar imaging demands wideband signals and large array apertures. Thus a vast amount of measurements is needed for a detailed reconstruction of the scene of interest. For practical through-the-wall radar imaging systems it is imperative to reduce the number of samples to cut down on hardware cost and/or acquisition time. Conventional image formation algorithms, such as delay and sum beamforming, generally need the full data set to obtain satisfactory images. However, compressive sensing (CS) can be used to reconstruct high-fidelity images from only a fraction of the measured data

Feature Extraction, Segmentation and Classification

Automatic detection of humans and objects of interest, e.g., concealed weapons or explosive material, is of high practical interest which is fundamental to follow-on tasks of target classification and tracking, image interpretation and understanding. However, it has been shown that the image statistics may vary dramatically in time and space. A practical detector, thus, has to adapt itself to changing and unknown statistics. Our research is focused on the design of fully automated detectors for Through-the-Wall radar imaging which require no prior knowledge and adapt to varying statistics.

Experimental setup
Automatically obtained 3D detection result

The aim of segmentation is to divide the total number of voxels in a 3D radar image into different classes, each of them representing e.g. targets of interest, noise, clutter, etc.
Segmentation is performed using a segmentation by classification scheme, where polaremtric features are used to generate homogeneous entities called Super-Pixels using Quick Shift technique, then these super-pixels are classified to its corresponding classes using Random Forests classifiers.

Segmentation by Classification scheme

Recent Results

The problem of target segmentation and classification in the image-domain with application to Through-the-wall radar imaging was addressed. We overcame the limitations of algorithm based on a pixel-grid which is not a natural representation of visual scenes. Perceptually meaningful entities obtained from a low-level grouping process, so called super-pixels are

considered. Further, simple geometrical and statistical descriptors that have

been used as features in the past are extended to polarimetric descriptors which make use of the whole polarimetric information in radar images. A framework of polarimetric feature extraction, over-segmentation (super-pixels), clustering, and classification is presented. An expandable sequential classifier based on random forests has been proposed to discriminate

targets from clutter returns and to provide further information about the discriminated targets. The experimental results demonstrate the usefulness of the proposed methods as desired target returns are discriminable from clutter returns and a further classification about target type and its nature is achieved.

One target Classification Result
2 Target Classification Result

Key Research Area

Electrical Engineering, Statistical Signal and Image processing

Supervisors

Prof. Dr.-Ing. A. Zoubir  (Signal Processing)
Prof. Dr. rer. nat. W. Stannat  (Probability Theory and Stochastic Analysis)

Contact

Ahmed A. Mostafa
M.Sc.

Address:

Dolivostraße 15

D-64293 Darmstadt

Germany

Phone:

+49 6151 16 - 24401 or 24402

Fax:

+49 6151 16 - 24404

Office:

S4|10-305

Email:

amostafa (at) gsc.tu...

 Print |  Impressum |  Sitemap |  Search |  Contact |  Privacy Policy
zum Seitenanfangzum Seitenanfang