An Approach of Data Fusion for FuzzyART Based Visual Recognition

Abstract

Visual Place recognition in robotics is motivated by the challenges of navigation and mapping. Given an image, a robot can map or recognize his environment. The ability to perform well is related to the quality of the information acquired and processed by the system. As the variation of the real world affects significantly the results, it can lead, in some cases, to a poor performance. The improvement is related to changes due to illumination, weather or season at outdoor with abundant features and textures for place recognition as in indoor environment, the robot will have less features or textures to process. In this paper, we present an approach of place recognition inspired by the FuzzyART. The input will be a Normalized Scanline Profile (NSP) and a pre-treatment will be performed by two fuzzy ART subsystems. The matching being the threshold for decision, we will fuse the matching output of the subsystems using a Sugeno-Takagi Fuzzy Inference System. The developed approach was tested on two different dataset to evaluate its ability to perform on outdoor and indoor environments and to limit the use of data storage. The results indicate that correct setting leads to optimal results.