Researches On Signal Processing And Target Recognition Of Ground Penetrating Radar For Subsurface Detection

Landmine is one of the conventional defensive weapons, buried in the ground ordeployed on the surface of the earth. Once it is triggered, it will hurt or kill the targetby explosion. As landmines are cheap and have remarkable e?ect, they are utilizedby many nations and organizations. However, landmines are easy to lay but hard toclear. The remained unexploded landmines have posed serious threat to civilians andrestricted the local economic development after wars.Ground Penetrating Radar (GPR) has played a great role in international humani-tarian demining. GPR is a non-contact detecting instrument, based on the variation ofthe underground electromagnetic properties. It detects the underground environmentby emitting high-frequency electromagnetic waves and receiving its echoes. Unlikethe air surveillance radar, it confronts the dispersive, anisotropic, strong-damped anddiversified underground lossy dielectric. The minimum wave length of the GPR ra-diowave should less than or equal to the maximum physical size of the buried object,otherwize, the radiowave can not be re?ected by the object. The history of GPR canbe traced back to the 1900s, most of the early GPRs are used for middle or deep levelunderground detection, but they can not help to detect subsurface small objects within10cm length and less than 0.5 meters deep. The detection of subsurface small objectsis usually a near-field task, requiring high-level signal resolution and fast computingspeed. GPR using Ultra Wide Band (UWB) technique can meet this need by its shorttransmitting time interval and wide range of frequency spectrum. The UWB GPRtechnology have been developed since 1990s and have made it possible for subsurfacesmall object detection.This paper comes from the research of subsurface anti-personal landmine detec-tion using UWB GPR. We utilize the data acquired by UWB Impluse GPR. We mainlyfocus on the signal processing technology of GPR including signal preprocessing, tar-get detection and location, feature extraction and target recognition. The contents and innovations of this paper are as follows:1. A clutter reduction method for GPR data is proposed based on Principal Com-ponent Analysis (PCA) and Adaptive Two-Sided Linear Prediction. First, theB-scan image is reconstructed using the subspace of secondary components ofPCA; then, we made the linear prediction using two reference scans before andafter the current scan; and last, Recursive Least Square is applied to employ thecorrelation of neighbor A-scan. As the proposed methed combines the overal-l information of B-scan, the reference information of anteroposterior A-scansand the recursive result of the neighboring A-scan, it gives best clutter reductionperformance in the comparative experiment.2. The detection and location of target hyperbolae are studied from two pointof views: the region-based description and the line-based description. Inregion-based target detection, we utilize multi-orientation and multi-scale Gaborwavelet filters to deal with the uprise, horizontal and descend edge of hyperbolae.The energy ratios of di?erent component images after filtering are calculated tofind the target location. In the line-based target detection, we proposed a com-pound method using image morphology and other image processing techniquesto detect perturbed hyperbolae. On the basis of that, the parameters of hyperbolaeare also fitted using Direct Least Square Fitting.3. Feature extraction of GPR data has been discussed through time-frequency anal-ysis of A-scans and texture analysis for the 2-D time-frequency spectrum. Westudied four di?erent time-frequency techniques for GPR A-scans, utilized fiveimage texture descriptors to extract the feature of the 2-D signature, and then e-valuated their discriminating ability using the normalized distance through FisherLinear Discriminant. The experiment shows that, the S-Transform provides thebest discrimination result, followed by Wigner-Ville Distribution, the Short-TimeFourier Transform and the Short-Time Hartley Transform performs the worst. Inorder to construct the feature vectors, we examined 20 texture feature descriptorsbased on Gray Level Occurrence Matrix and Texture Feature Coding Method, s-tudied their target detecting ability and mutual relationship. 4. A target recognition technique is also proposed. In the process of Ground Pene-trating Radar (GPR) probe, the distance between GPR antenna and buried targetsvaries from far to near, then from near to far. This process is re?ected by thechanging A-scan sequence of the received data. So a target recognition methodis proposed based on this variation process, which is modeled and recognizedusing one-way non-stride continuous Hidden Markov Model (HMM). We havediscussed the di?erent discriminating ability between models with di?erent num-ber of states and Gaussian Mixtures. Also, it is shown that the proposed HMMmethod based on the variation of A-scan sequences performs better than SupportVector Machine, which only relies on the unordered sets of A-scan samples.