| Summary: | Object localization and its pose estimation in a scene is an extremely important research area in computer vision. In fact, it forms an important first step for performing further manipulation tasks. Several techniques have been proposed in the literature that strive to capture this information. Shape matching is one such technique, as even simple hand drawn contours without the aid of other cues such as color, texture etc. can be used to delineate, identify, localize and compute object pose. In this work, the power and the flexibility of object centered and viewpoint dependent representations of objects from cognitive science are combined to estimate the 2D parameters of baggage in a captured image. Object-centered theories allow us to construe the problem of object representation in terms of its parts. In particular, the recognition by component model represents objects in terms of part primitives referred to as geons. The objects are parsed at the points of negative minima of curvature in accordance with the principle of transversality. This concept of object divisibility into parts can be exploited to extract the most visually salient part(s) for pose estimation. In this research, considering that the baggage is a two-part structure comprising of the salient body and the handle, a trained single layer feedforward neural network based on Extreme Learning Machine (ELM) and Harris corner points is used to extract the cuboidal contour of the baggage at fixed scale. The extracted cuboidal contour is then matched with stored cuboidal templates of different poses and aspect ratios using the Chamfer matching technique. This allows for obtaining the best template and the corresponding translation parameters thereby pointing to the integrated approach of object-centered and viewpoint dependent representations for baggage localization and pose estimation. This approach also aligns well with the concept of part saliency, selective attention to and processing of parts from the field of cognitive science in which the top-down influence of the task to be performed (pose estimation) is also taken into consideration. In this research, baggage in uncluttered scenes are considered for the purpose of 2D pose estimation based on which an algorithm is proposed. The success of the algorithm is demonstrated through simulations in MATLAB. Future research in this direction will enable a potential robotics solution for automated baggage handling for which pose estimation is a must.
|
|---|