Features that have been employed in prior object-recognition algorithms include points, edges, and textured regions. Each of these types of features offers different advantages and disadvantages in different applications. Line segments were chosen as the features to use in this algorithm because, relative to other features, they are better suited for recognition of partially occluded objects amid clutter.

The algorithm incorporates the assumption that at least one line of a model of an object that one seeks to recognize is detected as an unfragmented line in the image. As used here, “unfragmented” signifies that the affected image line is extracted from the image as a single continuous segment between the two end points of the projected model line. To be unfragmented, the line must also be unoccluded. Additional model lines must be present in the image for verification, but these may be partially occluded or fragmented.

The object-recognition process of this algorithm comprises three stages, following an approach that enables rapid focusing of computational resources on the highest-payoff hypothesized poses, thereby enabling a large reduction in the amount of computation time. In the first stage, approximate model poses are hypothesized and listed. Every pairing of a model line to an image line first contributes a pose hypothesis consisting of a similarity transformation. When a pair of non-parallel model lines and corresponding pairs of image lines form corner-like structures and the angles of the corners are sufficiently similar (within 45°), a pose hypothesis consisting of an affine transformation is added to the list of hypotheses. Typically, each model-to-image line correspondence contributes between one and six hypotheses to the list.