Probabilistic fusion of multiple algorithms for object recognition at information level

Abstract

Reliable object recognition is a mandatory prerequisite for service robots that operate in everyday environments. Typical approaches run a single classifier for the purpose of object recognition. However, no single algorithm proved to classify across all types of objects. We propose an approach that combines the recognition result of several methods working on different features. This reduces the effort and complexity of a single algorithm to recognize all known objects and makes the overall recognition robust. Known algorithms are extended to use a semantic output of a recognition probability for easy integration. To overcome the limitation of an algorithm to a class of objects based on their features, we introduce a probabilistic quality that defines how well an algorithm can recognize a known object type. The algorithms results are integrated using probabilistic methods to formulate a final belief. The approach is demonstrated in practical experiments in which a service robot recognizes and grasps similar appearing objects. The experiments show that the recognition is improved by probabilistic fusion of multiple algorithms.