Modeling and Information Processing
ATR - Automatic target recognition
Target recognition by SAR imaging
Measurements of the similarity between real SAR images and SAR images simulated from 3D models.
Due to its lateral and all weather view taking capabilities, SAR is an image-taking sensor particularly suited to the problem of battlefield surveillance, especially to meet a need for urgent programming. Within the context of photo-interpretation support activities, we have worked on shape recognition support algorithms for this type of imaging based, in particular, on the contribution of similarity measurements between a real SAR image and a simulated image. The images used are MSTAR data, Moving and Stationary Target Acquisition and Recognition, in the X band, resolution of 30 cm . This data concerns different objects of interest.
The proposed procedure is shown in figure 1. Using a candidate 3-D target model, we simulate a series of SAR images taken at different orientation angles using the view taking conditions of the real image to which we wish to compare them. These simulations are purely geometric. The images have three grayscale levels corresponding to the target, the shadow and the background. In parallel, the SAR image studied is classified in three categories (target, background and shadow) on the basis of the algorithm described in [2]. Then, a measurement of the correlation between the classified real image and the simulated image is used to determine the most probable orientation of the target for a given model. This processing sequence is repeated using the different models available. The correlation score for each of the models then gives an indication about the most probable nature of the target.
Figure 1 - Processing flowchart.
Figures 2 and 3 illustrate the processing in the case of a series of T72 images. Figure 2 on the left represents the estimation error in orientation for the T72 images when presented with the suitable model. It seems that these errors are quite low. The correlation values obtained for the best orientation between the series of T72 images and different model (T 72, leclerc tank, peugeot P4, mirage III) are plotted on the right in figure 2. We find that the T-72 and Leclerc tank curves are almost interlocked. Conversely, the Peugeot P4 and Mirage III curves are clearly below that of the T-72. We can deduce from these results that this method can distinguish a tank from a jeep or an aircraft but cannot differentiate between two types of tank.
Figure 2a - Estimated error in orientation depending on the real orientation of the view taken
for the T72 tanks in the MSTAR database.
Figure 2b - Example of model image superimposition for a T72 from the MSTAR database
and different target models.
Figure 3 - Example of superimposition of a T 72 image on different models
These first results show that, in the cases studied, a measure of similarity between a real image and a simulated image allows us to determine the target's orientation in relation to the carrier and eliminate some candidate models. Now, we need to continue with the experiments in order to determine this method's field of application more accurately.
References
[1] L. Fouque, Sur une méthodologie générale de fusion de Dempster-Shafer pour la classification markovienne d'images multirésolutions, mémoire de Thèse de l'INT, 2002
[2] E. Simonetto, Extraction 3D de structures industrielles sur des images RAMSES haute-résolution par radargrammétrie, mémoire de Thèse de de l'Université de Rennes, avril 2002,
[3] C. Lallemant, Mesures de ressemblance entre des images SAR réelles et des images SAR simulées à partir de modèles 3D , Rapport de stage Onera, septembre 2002.
