In this talk, I will review our recent work on model-based tracking and object recognition using range data, with a particular focus on algorithms designed to handle very sparse data comprising only thousands or hundreds of points. I will introduce a new algorithm based on an exploration of the ICP minimum potential well space. This algorithm explicitly maps out local minima in the well space, and embeds the location of these minima in a feature vector. The minima location have been shown to be very distinctive for different views of an object, and are used to index a database for recognition.

The algorithm has been shown to be both efficient and accurate, with recognition rates in the high 90% for segmented objects. The talk will conclude with some thoughts on the implications for the design of future sensors.

Les séminaires du LVSN ont habituellement lieu le vendredi à 11h30 dans la salle PLT-3775.