Improving Tracking by Integrating Reliability of Multiple Sources Luca Marchetti * , Diana Nobili † , Luca Iocchi ‡ Department of Systems and Computer Science University of Roma, “Sapienza” via Ariosto, 25 00185 Rome, Italy Email: * lmarchetti@dis.uniroma1.it † diananob@libero.it ‡ iocchi@dis.uniroma1.it Abstract—Tracking algorithms are often designed around optimistic assumptions on uncertainty model. Handling with conflicting data, however, requires specific strategies, that con- sider quality of information sources. To improve performance of tracking systems, the use of reliability, as evaluation of quality of data sources, has been proved to be a promising technique. In this paper we show how to use reliability of information sources to increase performance of tracking methods, using two different strategies: discount and pruning. We apply those two strategies in two different scenarios: landmark based mobile robot localization using Extended Kalman Filter and multi-agent object-tracking using Particle Filter. Experimental results show effectiveness of proposed methodology. Keywords: Multi-agent tracking, localization, Kalman fil- tering, Particle filtering, estimation, reliability. I. I NTRODUCTION One of the most important abilities of an autonomous agent is gathering information from the surrounding environment in order to build a symbolic representation of it. In particular, in many applications it is useful to track moving objects over time. This goal is usually achieved by tracking algorithms. Tracking algorithms can be divided in two classes: object tracking algorithms and self tracking, or localization, algo- rithms. The main goal of the former class is to locate the absolute position of moving objects over time. The latter one, instead, uses fixed objects as reference to locate position of a mobile base with on-board sensors. In this paper we will treat these two classes in the same framework. Therefore a general approach will be presented, and applied to multi-source object-tracking and localization problems. Despite this paper has precise focus on tracking methods, the same concepts can be applied to a generic data fusion system. A multi-source tracking architecture uses data coming from various input sources, “a priori” knowledge about the environ- ment and a known model of uncertainty. Therefore, the success of information fusion depends on how well the knowledge produced by the fusion process represents reality, which depends on how adequate data are, how good and adequate is the uncertainty model, and how accurate, appropriate or applicable is the prior knowledge. At a first step, the performance of a fusion process can be estimated by the error between real state and estimated one. This is an idealistic assumption, because it is almost impossible to relate true state with the estimated one. However, perfect observations are typically not available, due to low resolution of sensors, random noise in sensed environment, or because the process to be estimated cannot be well mod- eled. Consider for example observations extracted from image processing routines for a camera mounted on a mobile robot. The first level of uncertainty knowledge is usually provided by covariance matrices, but this is not sufficient to guarantee good performance. In fact, it is impossible to detect defective sources, or situation wherein the measurements gathered con- tain errors. In such situations it is thus necessary to develop a second order of fusion that involves reasoning about the conditions wherein measurements are taken. This introduces a higher level of knowledge, in the form of uncertainty of sources. Many authors proposed to deal with the second level of knowledge, by introducing the concept of reliability [1]. The definition of reliability involves two aspects: • observation reliability, related to single measurement of a sensor or information source; • source reliability, related to the quality of a whole source. In this paper we use reliability of information sources to improve tracking methods for mobile robotic platforms. The main idea is to define reliability of a source, as the result of processing information about the source itself. From this perspective, reliability is not associated with fusion results, but it can be viewed as an “a priori” knowledge about the operating situations. In the following sections, we will present a brief introduc- tion of the problem, showing an overview of current methods present in literature. Then we will formalize the problem of tracking within a reliability framework, focusing attention on two strategies available for reliability handling: discount and pruning. To evaluate effectiveness of our proposal we will present experiments in two different scenarios: landmark-based mobile robot localization and multi-agent object tracking. II. RELATED WORK The key point of successful tracking algorithms resides in the correct estimation of the system state. This is dependent on adequate parameters chosen per environment settings, correct sensor noise model and accuracy of sensors. The process of modeling beliefs has always some limitations, and models are valid only within a certain range. So, when combining information provided by many sources, we have to take into 1101