Estimating the number of people in a given area and knowing their positions open up numerous possibilities for a variety of smart data driven applications. Most of the existing systems either require active participation from people in the crowd or are too expensive to be deployed. The exponentially increasing adoption of smartphones by people and the ubiquitous Wi-Fi infrastructure motivated us to tackle this problem in a non-intrusive manner. This thesis focuses on designing a system to infer the crowd distribution in large indoor spaces. As we choose to be non-intrusive in approach, we incorporate low cost Wi-Fi sniffers into smart bulbs that are part of the lighting grids in buildings. The Wi-Fi data gathered from these smart bulbs is analyzed to obtain people count and the people’s position within a given area. Our aim is to estimate people’s location within 4x4 m2 grids with minimal number of Wi-Fi frames. A number of filtering and post processing mechanisms are proposed in order to eliminate false positives and to accurately identify the number of people within a given area. Extensive experiments are conducted in a realworld testbed with controlled settings as well as in test setups with no control (auditorium and a coffee corner). The improvised counting algorithm comes close to 75% of the ground truth. We adopt range free localization algorithms to estimate the position of people and evaluate these algorithms extensively. We propose enhancements on these algorithms to refine the position estimation accuracy and reduce the execution time. Range free localization algorithms were able to estimate the position 40% of the time within 2m when sniffers are placed 6m apart in a grid topology in a highly multipath testbed environment. Simulations indicate this number can increase up to 76% when sniffers are placed 4m for the same topology. Since the bulbs with sniffers cost more we minimize the overall deployment cost of the system by reducing the number of sniffers while maintaining an accuracy within a 4x4 m2 grid majority of the time. Extensive simulations are run with different topologies and sniffer densities in order to find this sweet spot.
iv