I explore the use of a wireless sensor network to automatically track materials in a warehouse, in order to comply with spatial constraints. This system, which is required to operate in real-time, employs sensor motes that use a combination of ultrasound bursts (chirps) and radio frequency to compute pairwise distance estimates. However, these estimates are based on data that is inherently noisy and unreliable. To address this issue, I outline a probabilistic algorithm that makes use of an explicit model of the ultrasound sensitivity pattern of Cricket motes. This approach makes use of particle filtering, a non-parametric method that is able to estimate the positions of potentially mobile motes. Particle filtering employs an arbitrary probability distribution as a model of the observed physical evidence, and uses sampling to produce estimated object states. I develop a number of potential observation models and compare their effectiveness with simulations based on collected real-world sensor data and computed distance measurements. These experiments suggest that the use of this algorithm coupled with probabilistic sensor models can accurately localize up to eight mobile motes to within an average of 8.5 cm, while updating their estimated positions based on collected distance estimates as often as once every second.





Download Full History