Adaptive Search Tree Database Indexing for Hand Tracking

Adaptive Search Tree Database Indexing for Hand Tracking
Nils Petersen, Didier Stricker
Computer Graphics, Visualization, Computer Vision and Image Processing 2012 IADIS International Conference Computer Graphics, Visualization, Computer Vision and Image Processing (CGVCVIP-2012), July 21-23, Lisbon, Portugal

Abstract:
Template matching is a powerful means for recognizing objects that are either complex to model or possess too few robust local features. The problem of tracking bare hands in an image sequence exhibits both of these properties. Similarity search within a database of precomputed hand views using template matching is therefore often used for hand detection and tracking. The challenge is the high-dimensional hand configuration space. Thus, a huge amount of templates is needed for accurate matching, making search speed a principal problem. In this paper we present an approach that can track hands at interactive frame rates within a database with more than four million hand views. Our main contribution is a method to quickly establish locally optimal search-trees to perform beam-searches within this database. This vastly reduces the time needed for a search run and thus affords the application of database indexing techniques in classical real-time tracking frameworks. We demonstrate the feasibility of our approach on synthetic and real world examples.

Adaptive Search Tree Database Indexing for Hand Tracking

Adaptive Search Tree Database Indexing for Hand Tracking
(Hrsg.)
Computer Graphics, Visualization, Computer Vision and Image Processing 2012 IADIS International Conference Computer Graphics, Visualization, Computer Vision and Image Processing (CGVCVIP-2012), July 21-23, Lisbon, Portugal

Abstract:
Template matching is a powerful means for recognizing objects that are either complex to model or possess too few robust local features. The problem of tracking bare hands in an image sequence exhibits both of these properties. Similarity search within a database of precomputed hand views using template matching is therefore often used for hand detection and tracking. The challenge is the high-dimensional hand configuration space. Thus, a huge amount of templates is needed for accurate matching, making search speed a principal problem. In this paper we present an approach that can track hands at interactive frame rates within a database with more than four million hand views. Our main contribution is a method to quickly establish locally optimal search-trees to perform beam-searches within this database. This vastly reduces the time needed for a search run and thus affords the application of database indexing techniques in classical real-time tracking frameworks. We demonstrate the feasibility of our approach on synthetic and real world examples.