Kernel-based particle filtering for indoor tracking in WLANs

• Authors: Victoria Ying Zhang;Albert Kai-Sun Wong
• Venue: Journal of Network and Computer Applications
• Year: 2012

Indoor localization using signal strength in wireless local area networks (WLANs) is becoming increasingly prevalent in today's pervasive computing applications. In this paper, we propose an indoor tracking algorithm under the Bayesian filtering and machine learning framework. The main idea is to apply a graph-based particle filter to track a person's location on an indoor floor map, and to utilize the machine learning method to approximate the likelihood of an observation at various locations based on the calibration data. Nadaraya-Watson kernel regression is adopted to interpolate the Received Signal Strength (RSS) distribution for nonsurvey points. The success of the proposed kernel-based particle filter (KBPF) lies in the fact that KBPF incorporates the environmental and motion constraints into the model and restricts particles to propagate on the graph which precludes the locations that the person is unlikely to be at, and that the developed nonlinear interpolation method is effective in inferring the RSS distribution for the non-survey location points which makes it possible to reduce the total number of survey locations. In addition, missing value problem is addressed in this paper, and different methods are compared through experiments. We conducted a series of experiments in a typical office environment. Results show that KBPF achieves superior performance than other existing algorithms. It even yields higher accuracy with only a small fraction of training data than others with a full training data set. As a consequence, by applying KBPF, sub-meter accuracy can be obtained while extensive calibration effort can be greatly reduced. Although KBPF is more computationally complex, it can still provide real time estimates.