A detailed description of the CLEARFAST system for thermal IR stand-off minefield survey is given. The system allows (i) a stand-off diurnal observation of hazardous area, (ii) detecting anomalies, i.e. locating and searching for targets which are thermally and spectrally distinct from their surroundings, (iii) estimating the physical parameters, i.e. depth and thermal diffusivity, of the detected anomalies, and (iv) providing panoramic (mosaic) images indicating the locations of suspect objects and known markers. The CLEARFAST demonstrator has been successfully deployed and operated, in November 2004, in a real minefield within the United Nations Buffer Zone in Cyprus. The paper describes the main principles of the system and illustrates the processing chain on a set of real minefield images, together with qualitative and quantitative results.
This paper deals with a three-dimensional thermal model for landmine detection problems and an inverse problem for reconstructing the physical parameters of buried objects. Moreover, solutions are given for the estimation of the soil thermal diffusivity and meteorological parameters, needed for solving the inverse problem. The paper describes the main fundamental principles of thermal modelling for buried object identification and illustrates the results on data acquired from a real minefield, together with qualitative and quantitative results illustrating the validity of the model.
This paper gives a comparison of two vehicle-mounted infrared
systems for landmine detection. The first system is a down-ward looking standard infrared camera using processing methods developed within the EU project LOTUS. The second system is using a forward-looking polarimetric infrared camera. Feature-based classification is used for this system. With these systems data have been acquired simultaneously of different test lanes from a moving platform. The performance of each system is evaluated using a leave-one-out method. On the training set the polarimetric infrared system performs better especially for low false alarm rates. On the independent evaluation set the differences are much smaller. On the ferruginous soil test lane the down-ward looking system performs better at certain points whereas on the grass test lane the forward-looking system performs better at certain points.
Polarimetric scattering models are developed to predict the detectability of surface-laid landmines. A specular polarimetric model works well only under the condition that there is either no sunlight or the sun is not close to the specular reflection direction. Moreover, this model does not give insight why certain man-made objects like landmines give a higher polarimetric signature than natural background. By introducing a polarimetric bidirectional reflectance distribution function (BRDF) the specular model is extended. This new model gives a better prediction of the polarimetric signature and gives a close match to the measurements of landmines with different casings as well as the sand background. The model parameters indicate that the landmines have a lower surface roughness and a higher refractive index, which is the reason why these objects are detectable from the background based on their polarimetric signature.
Several sensor fusion approaches are in use for detection of land mines. These are based on different techniques and show different results. Meaningful comparisons of fusion algorithm performance are difficult, because the performance achievable may be limited by the data acquired and the sensors used. Especially in comparing such dissimilar situations, it is important to compare measures that are not dependent of sensor peculiarities or layout of test lanes. Nonetheless, a quantitative comparison of algorithms is necessary to identify the most effective fusion techniques. This comparison must sometimes be done when the algorithms are being applied to different data sets. In this paper we describe a methodology to evaluate the performance of sensor-fused mine-detection systems. This process can be used to compare different mine-detecting sensors in well- defined scenarios.
High detection performance is required for an operational system for the detection of landmines. Humanitarian de-mining scenarios, combined with inherent difficulties of detecting landmines on an operational (vibration, motion, atmosphere) as well as a scenario level (clutter, soil type, terrain), result in high levels of false alarms for most sensors. To distinguish a landmine from background clutter one or more discriminating object features have to be found. The research described here focuses on finding and evaluating one or more features to distinguish disk-shaped landmines from background clutter in infrared images. These images were taken under controlled conditions, with homogenous soil types. Two methods are considered to acquire shape-based features in the infrared imagery. The first method uses a variation of the Hough transformation to find circular shaped objects. The second method uses the tophat filter with a disk-shaped structuring element. Furthermore, Mahalanobis and Fisher based classifiers are used to combine these features.
A specular model has been used to predict the passive polarimetric infrared (IR) signature of surface-laid landmines. The signature depends on the temperature of the landmine and the sky radiance. The temperature of the landmine is measured using a thermocouple. The signature itself is measured using a polarimetric IR camera setup. The predictions are fit to the measurements using the refractive index as an optimization parameter. The obtained refractive indices of each landmine type are consistent, but for the PMN landmine much lower than determined in a previous indoor experiment. Throughout the measurement day, the average landmine polarimetric signature was higher than the average background signature. Moreover the polarimetric signature appears to be a more robust indicator of the shape of the landmine's top surface than the normal IR signature. A simulator of passive polarimetric imagery is also being developed. That work is based on a physical model for both the thermal and radiometric processes, and it includes a finite-element solution for the heat transfer problem, ray tracing to describe the incident sunlight and the effects of shadowing, and analytical models for the Mueller matrices of rough dielectric surfaces. Preliminary results from that model show substantial qualitative agreement with measured images.
Linear polarization of Thermal InfraRed (TIR) radiation occurs whenever radiation is reflected or emitted from a smooth surface (such as the top of a landmine) and observed from a grazing angle. The background (soil and vegetation) is generally much rougher and therefore has less pronounced linear polarized radiation. This difference in polarization can be used to enhanced detection of land mines using TIR cameras. A measurement setup is constructed for measurement of polarized TIR images. This setup contains a rotating polarization filter which rotates synchronously with the frame sync of the camera. Either a Long wave InfraRed (LWIR) or a Mid Wave InfaRed (MWIR) camera can be mounted behind the rotating polarization filter. The synchronisation allows a sequence of images to be taken with a predefined constant angle of rotation between the images. Out of this image sequence three independent Stokes images are calculated, consisting of the unpolarized part, the vertical/horizontal polarizations and the two diagonal polarizations. An initial model is developed that describes the polarization due to reflection of and emission from a smooth surface. This model predicts the linear polarization for a landmine `illuminated' by a source that is either hotter or cooler than the surface of the landmine. The measurement setup is used indoors to validate the model. The measurements agree well with the model predictions.
In this paper we introduce the concept of depth fusion for anti-personnel landmine detection. Depth fusion is an extension of common sensor-fusion techniques for landmine detection. The difference lies within the fact that fusion of sensor data is performed in different physical depth layers. In order to do so, it requires a sensor that provides depth information for object detections. Our ground-penetrating radar fulfills this requirement. Depth fusion is then taken as the combination of the output of sensor fusion of all layers. The underlying idea is that sensor fusion for the surface layer has a different weighing of the sensors when compared with the sensor fusion in the deep layers because of apparent sensor characteristics. For example, a thermal IR sensor hardly adds information to the sensor fusion in the deep layers. Furthermore, GPR has difficulties suppressing clutter in the surface layer. As such, the surface fusion should emphasize on the TIR sensor, whereas sensor fusion in the deep layers should have a higher weighing of the GPR. This a priori information can be made explicit by choosing for a depth-fusion approach. Experimental results form measurements at the TNO-FEL test facility are presented that validate our depth-fusion concepts.
The infrared (IR) radiation emitted or reflected in an off- normal direction from a smooth surface is partially polarized. This principle can be used for enhanced discrimination of targets from backgrounds in a marine environment. It has been shown that (man-made) targets do not demonstrate a pronounced polarization effect when observed from near normal direction whereas the sea background radiation has a certain degree of polarization in slant observation path. A measurement setup has been constructed for collecting polarized IR imagery. This setup contains a rotating polarization filter that rotates synchronously with the frame sync of the camera. Either a long wave IR (LWIR) or a mid wave IR (MWIR) camera can be mounted behind the rotating polarization filter. The synchronization allows a sequence of images to be taken with a predefined constant angle of rotation between the images. Out of this image sequence three independent Stokes images are constructed, containing the normal intensity part, the vertical/horizontal polarization and the diagonal polarization. Up to 20 full linearly polarized images can be acquired per second. Measurements are taken at the North Sea coast with this setup. The recorded images are analyzed to determine the influence of polarization on the detection of small targets in such an environment. Furthermore differences between polarization contrasts in MWIR are analyzed.
To acquire detection performance required for an operational system for the detection of anti-personnel landmines, it is necessary to use multiple sensor and sensor-fusion techniques. This paper describes five decision-level sensor- fusion techniques and their common optimization method. The performance of the sensor-fusion techniques is evaluated by means of Receiver Operator Characteristics curves. These techniques are tested on an outdoor test facility. Three of four test lanes of this facility are used as training set and the fourth is used as evaluation set. The detection performance of naive Bayes, Dempster-Shafer, voting and linear discriminant are very similar on both the training and the evaluation set. This is probably caused by the flexibility of the sensor-fusion techniques resulting into similar optimal solutions independent of the fusion technique.
In this paper the landmine detection performance of an IR and a visual light camera both equipped with a polarization filter are compared with the detection performance of these cameras without polarization filters. Sequences of images have been recorded with a rotating polarization filter in front of the cameras.
In this paper the multi sensor fusion results obtained within the European research project GEODE are presented. The layout of the test lane and the individual sensors used are described. The implementation of the SCOOP algorithm improves the ROC curves, as the false alarm surface and the number of false alarms both are taken into account. The confidence grids, as produced by the sensor manufacturers, of the sensors are used as input for the different sensor fusion methods implemented. The multisensor fusion methods implemented are Bayes, Dempster-Shafer, fuzzy probabilities and rules. The mapping of the confidence grids to the input parameters for fusion methods is an important step. Due to limited amount of the available data the entire test lane is used for training and evaluation. All four sensor fusion methods provide better detection results than the individual sensors.