Advanced GPR Data Processing Algorithms for Detection of Anti-Personnel Landmines Summary Ground Penetrating Radar (GPR) is seen as one of several promising technologies aimed to help mine detection. GPR is sensitive to any inhomogeneity in the ground. Therefore any APM regardless of the metal content can be detected. On the other hand, all the inhomogeneities, which do not represent mines, show up as a clutter in GPR images. Moreover, it is known that reflectivity of APM is often weaker than that of stones, pieces of shrapnel and barbed wire, etc. Altogether these factors cause GPR to produce unacceptably high false alarm rate whilst it reaches the 99.6% detection rate which is prescribed by an UN resolution as a standard for humanitarian demining. The main goal of the work presented in the thesis is reduction of the false alarm rate while keeping the 99.6% detection rate intact. To reach this goal a set of data processing algorithms is developed and organized into an unsupervised target detection scheme. These algorithms are dedicated to clutter suppression and simultaneous detection of APM signatures in both GPR raw data and images resulting from them. The developed algorithms constitute together the following achievements: ¿ An unsupervised generalized likelihood ratio test-based feature fusion framework; ¿ A waveform based target detection/clutter suppression; ¿ Advanced methods for construction of GPR maps The unsupervised generalized likelihood ratio test based feature fusion framework, which has been suggested in this thesis, takes as input an arbitrary amount of confidence maps corresponding to training and testing sites. The output of the framework is a list of target locations. The framework uses training data which can come from independent and non-coincident measurements with different radars and even sensors. The data from each of the sensors are processed independently to result in several detection lists. Every detection in these lists is accompanied with one or several features each represented by a scalar number. A decision level fusion is applied to reconcile the lists i.e. to associate the detections in them with the appropriate physical locations. Then the binary hypothesis testing is executed for the reconciled locations separating them on clutter and target lists. The generalized likelihood ratio test is employed to this end. The feature pre-normalization via Johnson¿s transform in suggested by the author to be used prior the testing. It is shown in the thesis that such approach outperforms the direct generalized likelihood ratio testing ad. hoc. based fusion techniques. The waveform based target detection/clutter suppression algorithm, which detects disc-shaped APM in heavy clutter with low false alarm rate, has been developed by the author. The algorithm detects a class of low-metal APM with a cylindrical shape (such as PMN2, M14, and NR22 etc.) using just a single reference target return. It suppresses clutter responses from friendly objects while marking the presence of targets with sharp monopulses and preserving the spatial pattern inherent to localized objects. The algorithm is insensitive to the reflectivity and physical diameter of the target and also tolerates certain volatility in the properties of the hosting soil. This algorithm is superimposed with a focusing technique to further improve the mine detectability. A number of improved projection techniques, which allow better detection of APM in focused GPR images is also developed by the author. These utilize the prior knowledge on the character of the spatial correlation properties of target images and allow detection of the burial depth of the target. The algorithms suggested in the thesis were tested on the data acquired during two separate measurement campaigns held at the special facilities for testing of mine detection systems. It has been shown, that the fused multi-feature detection that uses the algorithms reported in this thesis, significantly decreases the false alarm rate in comparison to the previously published studies for the same minefields.@en

Successful detection of antipersonnel landmines often requires decision making on the basis of more than one decisive feature. The fusion of the available features should be done in a statistically optimal way. Theoretically this might be achieved by Bayesian-based criteria. However application of Bayes' criteria in general form is computionally difficult and may lead to overtraining of classification algorithms. To avoid this situation linear and quadratic fusion rules are often used in practice. However these rules produce optimal results only when the decisive feautures are normally distributed, which is not always guaranteed in practice. To improve the performance of the linear and quadratic detectors we suggest a normalization of the decisive features by means of Johnson's transform prior to the fusion. We describe the normalization algorithm and demonstrate that it improves performance of the linear and quadratic classifiers. The performance is judged in terms of the ROC curves.@en

In this paper we propose a new algorithm for enhancement of imaging of plastic cased antipersonnel landmines using a video-impulse GPR. The algorithm is implemented as a non-linear waveform based signal processor integrated with a SAR focusing procedure. The algorithm constructs a SAR-like image of the subsurface using the ouputs of the signal processor searches for presence of reference waveform in raw ID GPR echo-returns and compresses in time all responses similar to it. Simultaneously, the responses with different waveforms which presumably correspond to clutter, are suppressed. The shape of the predefined reference waveform depends on the angle at which the imaged point is illuminated. This algorithm is integrated into an automated data processing and mine detection scheme generating a detection list. The reference waveform and other algorithm parameters are determined from training datasets acquired in a controlled environment. The performance of the sequence of the signal processor without angle dependency and SAR and against a scheme involving a cross-correlation of the input with the same reference wavelet. The improvement achieved by the suggested algorithm is demonstrated in terms of ROC curves.@en

In this paper we propose a new algorithm for enhancement of plastic cased antipersonel mines detection using a video-impulse GPR. The algortihm is implemented as a non-linear signal processor integrated with a SAR focusing procedure. The new algorithm constructs SAR-like image of the subsurface at each point summing the outputs of the signal processor. The signal processor searches for presence of reference waveform in raw !D GPR echo-returns and compresses in time all responses similar to it. Simultaneously, the responses with different waveforms, which presumably correspond to clutter, are suppressed. The shape of the predefined reference waveform depends on the angle at which the point being imaged is illuminated. This algorithm is integrated into an automated data processing and mine detection scheme generating a detection list. The reference waveform and other algorithm parameters are determined from training datasets acquireed in a controlled environment. The performance of the new algorithm is compared against the performance of the sequence of the signal processor without angle dependency and SAR and agianst a scheme involving a cross-correlation of the input with the same reference wavelet. The improvement achieved by the suggested algorithm is demonstrated in terms of RTOC curves.@en

None given@en

The automated detection of plastic-cased antipersonnel landmines in 3-D high-resolution UWB radar images is considered. The 3-D images are projected in different ways onto 2-D confiedence maps, in which the automated APM detection is made by marking local maxima in the maps followed by thresholding. Specifically, we consider the confidence maps of the following types: energy projection (EP), windowed energy projection (WEP) and introduction here alternating sign windowed energy projection (ASWEP). The latter is shown to allow diminishing of the new technique are described by means of receiver operator curves (ROC). The ROC curves are built for the datasets acquired over several minefield simulation sites under different environmental conditions.@en

The performance of GPR as an anti-personal mine (APM) detector heavily depends on the quality of the acquired data. A scheme of the quantitative evaluation of the quality of raw GPR data in terms of stability of certain parameters and the set of the procedures, which comppensate the instabilities found are presented. We demonstrate the performance of the data pre-processing chain (DPPC), whihc includes spectral filtering of A-Scans, control of the mutual alignment of B-scans, and time and voltage axis's instabilities compensation in the whole C-scan. The performance of the DPPC is judged both in terms of the stabilization of certain statistical parameters of the data, and by its effect on the Receiver Operation Characteristics (ROC) curve.@en

The results of the measurements campaign held during the autumn of 2004 in a sand pit located at the TNO-DS&S ( formerly known as TNO Physiscs and Electronics Laboratory (FEL) GPR test facility "Waalsdorp" in the Hague, The Netherlands, are presented. Some contaminiation of the sand in the pit with small scatterers (e.g. sea shells) comparable in size with smaller anti-personnel mines (APM) and roughness of the soil/air interface make the APM detection a non-trivial task. The performance of the detector relevant to the above mentioned data processing schemes and resulting images are compared and discussed. The performance of the detector relevant to the above mentioned data processing schemes are judged with receiver operating characteristics (ROC) curves built on the basis of the automated.@en

The present paper deals with feature-based landmine detection by means of UWB Ground Penetrating Radar (GPR). A waveform of target response is used as as discriminative feature for two feature-based detection schemes. The output of each DPS is a confidence map obtained by means of energy projection of a 3D distributed feature onto a 2D surface. An automated detection is made over the obtained confidence maps and mines detectability is analyzed on the basis of the Receiver Operation Characteristics curves. Performances of the feature- and energy-based data processing schemes (DPS) applied to the experimental data are compared. It is shown that suggested wave-form based detection is stable to change of the environment.@en