Fast and Robust Feature-based Recognition of Multiple Objects

Abstract

In robotics, one crucial requirement to a visual system is robust and efficient recognition of multiple objects. While in many available systems the focus is on tracking, the main problem still is to recognize objects in an arbitrary scene within a database of multiple objects. For any tracking system, recognition is needed for initialization and therefore always built in. However, the task of recognition becomes considerably harder, when learning and recognizing multiple objects. In this paper, we present a system, which accomplishes this task for textured objects robustly and efficiently. Our system is based on texture features, combining principal component analysis, k-means clustering and kd-tree search with best-bin-first strategy. We evaluated our system in several real-word scenarios, and present experimental results in a kitchen environment. Within a database of 20 objects, our system can analyze an arbitrary scene in less than 350 ms on a 3 GHz CPU