Abstract
We report a novel method of visual modeling and detecting 3D rigid objects randomly located in complex cluttered environments. Models are built (in reference scale) using visual saliencies (interest points) in template images presenting the object of interest from various viewpoints. By matching interest point detected in camera-captured images (relative scale is used there) to visual saliencies from the database, the target objects are detected and verified. If a stereovision system is available, the procedure is separately performed for both cameras. Subsequently, the corresponding matches from both images can be used to solve the fundamental correspondence problem. The paper briefly discusses methodology and presents exemplary experimental results in vision-based robotic aplications.