A computer vision system to detect 3-D rectangular solids

We present a computer vision system to detect 3D rectangular objects. We first describe the method to detect rectangular solids in real video/images in arbitrary orientations, positions, distances from the camera and lighting. The method works by detecting junctions and adjacent edges of rectangular solids. If a rough reference image of the background is available, that can also be used. We have tested our system on several hundreds of real images and video sequences. In particular, we evaluated the system performance by plotting receiver operating characteristic carves (probability of detection versus probability of false alarm). These curves were plotted for results on 500 images and video sequences acquired an a scene with rich background structure; that is, the scene had a large number of background lines and rectangles. In such an environment, we achieved 93% detection at a 13% false alarm rate. Potential applications of this system include detection of packing boxes, trailers of trucks and rectangular buildings. This system could be used for video indexing or for video surveillance in a security monitoring system.