Robust localization for mobile robot by K-L Divergence-based sensor data fusion

Abstract

This paper concerns the sensor data fusion method in mobile robot localization. Sensor data fusion methods using particle filter have been a major technique to estimate robots's state from noisy sensor observation. Generally, the sensors assumed to be in the nominal state of work. In realistic contexts, however, sensor characteristics may change online depending on sensor functioning conditions, and deteriorate estimation accuracy. To generate robustness to disturbance, faulty sensors should be identified and the fusion rule should be updated accordingly. These processes should be done online without access to ground truth, known as Fault Detection and Isolation (FDI) problem. The major approach to FDI problem in mobile robot localization has been the model-based online monitoring of the sensor state utilizing sensor fault models. However, comprehensive theoretical models for complex systems are difficult to obtain, and in some situation impossible to derive. Therefore, authors proposed a new sensor data fusion method for mobile robot localization adopting the idea of majority voting FDI techniques, which has been a proven technology in the field of safety critical control, utilizing the increasing and diversifying onboard sensors thanks to the technological development in recent years. Under the strong assumption of majority voting that the majority of similar signals provide the truth and any dissimilar signal is the result of a sensor fault, faulty probability distributions computed from sensor observations by a particle filter are detected and isolated utilizing K-L Divergence, a measurement for distance between two distributions in information theory. The proposed method is examined on a computer simulation, showing the possibility of generating robustness to the sensor data fusion in the mobile robot localization.