Spatio-temporal constraints for matching view-based descriptions of 3D objects

We propose a 3D object recognition method based on first extracting a compact description of image sequences, and then matching these descriptions with a two-steps strategy. The compactness of the description, made of a set of time-invariant local features, allows us to use a simple nearest neighbour matching to obtain an initial set of recognition hypothesis; spatio-temporal constraints help us to confirm or to reject these hypotheses. We carried out an extensive experimental analysis to assess our method against changes in illumination, abrupt scale changes, occlusions, clutter, view-point variations. Since our approach is based on the use of image sequences we also tested it with different camera motions. Our description strategy may be also applied to relatively simple poor-textured objects, since few stable features are enough for recognition.