Modeling and Information Processing
ATR - Automatic target recognition
Identifying objects described by bipoint statistics
Context
The objective of the reconnaissance function is to associate a signal measured (image, video sequence, etc.) with an indication of the nature of the object which produced that signal.
Here, we are interested in signals typical of those processed in remote sensing: images produced by an onboard sensor with known characteristics (resolution, transfer function, noise level, view-taking conditions, etc.). In the target contexts, the objects to be recognized are movable targets with varied contrasts and low resolutions. The only graphic indicators detectable with a certain degree of confidence are places with a high variation in local contrast - the object's contours. However, the resulting contour configurations are subject to noise, spatially folded and relocated (Figure 1). The objective of the work described below is to design a method for representing these contour configurations and a recognition principle adapted to their format.
Figure 1: examples of contours detected for the same orientation (effects of folding and varied contrasts)
Principles of the approach
The contour configurations are represented by families of bipoint statistics conditioned by the local gradients. Different conditioning values allow us to locate different parts of the objects observed (figure 2). Each image is then interpreted as an independent identically distributed (i.i.d.) controlled process. This method of shape analysis constitutes an invariant representation by rotation.
Figure 2 - Block diagram of the approach.
The recognition principle is to build a test sequence on the sequence of observations generated by a sampling of the bipoint characteristics. The final decision is the result of the accumulation of likeliness of each bipoint observed in relation to the types of objects that may be present in the image analyzed. The choice of the bipoints sampled is regulated by a steady control law that optimally combines the different conditioning. It can be demonstrated that the recognition error decreases exponentially towards zero as a function of the number of samples observed (large deviations theory).
Experiments
The algorithm was tested on a problem of identifying aircraft simply from their silhouette (Figure 4) and the distance to the object. The learning consisted of estimating the optimum combination of bipoint statistics for different interference parameters simulated from the silhouettes: folding, low affine distortion, contour sensing noise.
Figure 4 - Silhouettes of the objects to be recognized (Mirage IV, Rafale, Eurofighter, Mirage III, Alpha jet, Fouga and Jaguar).
After learning, the recognition stage consists of accumulating the likelinesses on a sample of bipoints.
Fouga [vidéo]
Eurofighter [vidéo]
References
[1] S. Herbin, Active Sampling Strategies for Multihyposthesis Testing, in EMMCVPR 2003, Lecture Notes in Computer Science, vol. 2683, Springer Verlag, pp. 97-112.
[1] S. Herbin, Robust Multihypothesis Discrimination of Controlled I.I.D. Processes, in ICPR04, vol I, pp 97-112.
