Ranking of features for classifying industrial objects

An algorithm for the recognition and localization of partially occluded objects is presented here. It is assumed that at least three corners, not necessarily consecutive corners, of all the objects present in the scene are visible. No restriction is made on the position and orientation of the object. For any particular object the position and rotation transformations are estimated by matching the triangles of the model and the scene. The ambiguity of the same triangle being present in more than one object model is resolved by a penalty function based on the area of mismatch. A new concept of feature ranking has been introduced so as to help the recognition algorithm in terms of within object variation as well as between object discriminability. It helps in reducing the number of initial hypothesis. A complete system has been designed and implemented and tested on a variety of scenes. The results clearly demonstrates the effectiveness of the proposed method.<
>