Classification of Wi-Fi Sensor Data for a Smarter City : Probabilistic Classification using Bayesian Statistics

Abstract: As cities are growing with an increasing number of residents, problems with the traffic such as congestion and larger emission arise. The city planners have challenges with making it as easy as possible for the residents to commute and in as large scale as possible to avoid vehicles. Before any improvements or reconstructions can be made, the traffic situation has to be mapped. The results from a probabilistic classification on Wi-Fi sensor data collected in an area in the southern part of Stockholm showed that some streets are more likely to be trafficked by cyclists than pedestrians while other streets showed the opposite. The goal of this thesis was to classify observations as either pedestrians or as cyclists. To do that, Bayesian statistics was applied to perform a classification. Results from a cluster analysis performed with K-means algorithm were used as prior information to a probabilistic classification model. To be able to validate the results from this unsupervised statistical learning problem, several model diagnostic methods were used. The final model passes all limits of what is considered to be a stable model and shows clear signs of convergence. The data was collected using Wi-Fi sensors which detect a device passing by when the device is searching the area for a network to connect to. This thesis will focus on data from three months. Using Wi-Fi sensors as a data collection method makes it possible to track a device. However, many manufacturers produce network interface controllers that generate randomized addresses when the device is connecting to a network, which makes it difficult to track the majority of the devices. Therefore, Wi-Fi sensor data could be seen as not suitable for this type of study. Hence it is suggested that other methods should be used in the future.