Specific object recognition using iso-luminal contours : a thesis presented in fulfilment of the requirements for the degree of Master of Engineering in Mechatronics at Massey University, Palmerston North, New Zealand
Object recognition is a broad topic in the study of computer vision. In this case the task of
distinguishing between specific instances of various objects is addressed. The ability to
perform this task would allow robots to operate in unstructured environments, allowing
greater and more efficient automation of many tasks. Techniques currently proposed tend
to have low accuracy rates, high processing time, or both. This research seeks to establish
a method that can quickly and accurately find instances of objects within a scene.
Iso-luminal contours were used to gather the initial data, from which higher level features
were extracted. Basic geometric features were used as the intermediate data, consisting of
lines, arcs, and lobes (a custom type suited to describe corners). The high level data was a
custom type, called blocks; each block contains a few features and describes the spatial
relationships between them. The features and blocks are designed to be spatially
invariant, so the blocks are directly compared to determine which objects are in a scene.
The objectives of this research were not met. The results show the geometric features
were not robust to changes in image sets, although they did work well with the image set
they were developed with. Unfortunately this means the performance of the subsequent
'block' related steps cannot be established. Most of the work was focussed on this aspect.
Future work would entail increasing the robustness of the features part of the algorithm,
and then gauging if the block based research is of practical use.
It is thought that the research results were poor because feature extraction was poor. It is
further thought that the high level analysis has merit.