Recognising 2-D shapes from incomplete boundaries

In many cases where object identification is necessary complete boundaries are not available due to poor contrast or occlusion. In such cases global shape descriptors are not effective. We describe a multiresolution hypothesis and verify a method for matching characteristic local segments of a boundary with pre-defined models. Boundary shape is characterised using a model of the distribution of dominant points within an image. A multiresolution pyramid is used to define, at the corner of each tile, anchor points for one end of virtual springs; the other end of each virtual spring is attached to a selected "dominant point" on the boundary. The virtual force at each corner is retained as a descriptor for that point in the pyramid. The dominant boundary points are selected by identifying points that exhibit significant curvature at a range of scales. The performance of this algorithm was evaluated using objects that are similar in appearance. With 20% of the boundary occluded 100% correct recognition was achieved and with 50% occlusion the correct recognition rate was 63%. Noise sensitivity was evaluated by the addition of boundary displacement noise with a standard deviation of up to /spl plusmn/5 pixels (/spl plusmn/5-10%). At this noise level the correct recognition rate fell from 100% to 93% and a greater depth of search (in resolution) was required to establish a high level of confidence in object recognition.