dressed because of the nature of data transmission. Radio frequency signals are signicantly impacted by changes in the environment, for example change in number of people and obstructions in the building, the relative position and direction of the user and ones machine, even temperature change can have an important eect. Proposed approach There is extensive research in the eld of robotics concerning localization techniques under uncertainty. Usually this uncertainty is caused by the the nature of the sensor readings, often ultra-sound proximity sensors, which is similar to the uncertainty in radio-frequency based signals. Mobile robot localization is the problem of determining a robots pose from sensor data. The problem in hand is similar to the problem that in the eld of robotics researchers are trying to solve with localization techniques. The algorithm that has proven to be the most promising one in the case of mobile robot position estimation is Monte Carlo Localization [3]. It is a stochastic approximation algorithm that results in a distribution over the possible locations of the tracked robot. Specically, in order to locate and track a mobile computer with the aim of a wireless network infrastructure one actually has to solve the global localization problem. The initial position of the machine must be determined from scratch and subsequently tracked. Initially the problem of self-localization should be addressed. In this context, the position is a relative vector according to a networks base station. The next step is to use a number of pursuers, now the targets position should be dened as relative position according to a pursuers machine. Although not apparent, this implies that the pursuer should be self-localized in regard to a base station. It is already obvious that the problem can have many dierent instances, ranging from the case of a xed target without any pursuers, to the case of a mobile target and xed pursuers and nally to the case of mobile target and pursuers. Step by step description of proposed implementation The nature of the overall objective leads to a step by step encounter with many of the problems that will come up during the implementation of this project: 2
Initially, necessary resources need to be allocated. This mainly refers to hardware equipment, such as laptop computers and network cards. The cards have to be installed and tested in terms of how they communicate with the wireless ethernet system. We also have to get familiar with various technical features in the current installation in Duncan Hall. The second step requires the alteration of the device drivers so as to get the signal strength from the network cards. Algis Rudys has kindly oered to assist our team in this step. It is very important to have early in this project a visual representation of the signal strength. That is why we intend to build in this step a tool that will help us visualize and understand the signal attributes in a better way. Then we will be able to experiment with sensor fusion. We will also have to test how the walls, the people and the relative orientation of the network card aects the signal strength. In the next step, a model will be built that will assist the localization technique. We have two dierent alternatives for this problem: We can place a laptop computer in various locations in the environment and measure the signal strength in every one of those positions. In this way, we would have constructed a map of expected signal strengths. On the other hand, we could create a model of the way the signal is propagated in space in terms of distance and how is eected by the presence of obstacles, such as walls. In this case, we could make use of a spatial map of the world, where we would know the location of base stations. Then we would apply our model of signal propagation in this world, since with the aim of the map we would know the number and the size of walls that a signal had to propagate if it was originated from a certain direction according to a base station. Our aim, though, is to achieve localizing the machine without having to make use of a spatial map. Usually an updated and accurate map of the environment is dicult to be found or created.
Another model that will assist the MCL algorithm is a model of motion of the machine. At this point, we will have to assume a bound on the speed of a user carrying a laptop computer. Thereafter, the Monte Carlo Localization algorithm should be implemented and tested along with the models we will have created. The objective is that self-localization and tracking should be supported at this level. The next step is to add pursuers. A pursuer should rst of all be selflocalized. Initially, we can consider the pursuers are xed and only the target is moving. If all the parts are correctly self-localized, then in order to get the useful relative direction of the target in regard to the pursuers it is just a matter of some simple transformations. Then it is interesting to test the eciency of the proposed technique when everybody is mobile, both the pursuers and the target Following, further improvements must take place. The latency will probably be a problem that must be confronted. Furthermore, Monte Carlo probably will need ne tuning, for example it would be interesting to see how it will work with fewer points. Changes in the models also may lead to better convergence.
References
[1] Paramvir Bahl and Venkata N. Padmanabhan: RADAR: An InBuilding RF-based User Location and Tracking System, Proc. IEEE INFOCOM, March 2000, http://www.research.microsoft.com/sn [2] Paramvir Bahl, Venkata N. Padmanabhan and Anand Balachandran: Enhancements to the RADAR User Location and Tracking System, Technical Report, MSR-TR-2000-12 [3] Sebastian Thrun, Dieter Fox, Wolfram Burgard and Frank Dellaert: Robust Monte Carlo Localization for Mobile Robots To appear in Articial Intelligence, 2001.