Object detection by optical correlator and Intelligent Recognition and Surveillance systems

We report our recent work on robust object detection in high-resolution aerial imagery in urban environment for Intelligence, Surveillance and Recognition (ISR) missions. Our approaches used the simple linear iterative clustering (SLIC) algorithm, which combines regional and edge information to form the superpixels. The irregularity in size and shape of the superpixels measured with the Hausdorff distance served to determine the salient regions in the very large-scale aerial images. Then, the car detection was performed with both the component-based approach and the feature-based approach. We merged the superpixels with the statistical region merging (SRM) algorithm. The regions were described by the radiometric, geometrical moments and shape features, and classified using the Support Vector Machine (SVM). The cast shadow were detected and removed by a radiometry based tricolor attenuation model (TAM). Detection of object parts is less sensitive to occlusion, rotation, and changes in scale, view angle and illumination than detection of the object as whole. The object parts were combined to the object according to their unique spatial relations with the deformable object model. On the other hand, we used the invariant scale invariant feature transform (SIFT) features to describe superpixels and classed them by the SVM as belong or not to the object. All along our recent work we can still trace the brilliant ideas by H. J. Caulfield and other pioneers in early days of optical pattern recognition for improving the discrimination of the matched spatial filter with linear combinations of cross-correlations, which have been inherited transformed and reinvented to achieve tremendous progress.