Research

CrowdMeter — Congestion Estimation in Train Stations

CrowdMeter is a participatory system that leverages sensor data collected from users' phones during their daily train commutes to gauge real-time congestion in railway stations. It tracks each passenger's position and identifies their context (waiting for a train, buying a ticket, etc.) along the trajectory from the entrance to the train. From these, novel features capture the passenger's behavior (walking pattern) and ambient environment (sound) that indicate congestion. The system highlights each area with a color (blue / yellow / red) for low / medium / high congestion. Field experiments at 29 different train stations in Japan show CrowdMeter infers congestion accurately, highlighting its promise as a ubiquitous travel-support service.

Core Body Temperature Estimation

We proposed a novel method to estimate core body temperature using a human thermal model that accounts for individual differences. The method employs Gagge's two-node model to simulate change of core and skin temperature by calculating heat production and heat exchange between the core node, the skin node, and the environment. In a warm-up phase, the optimal parameter set representing individual thermoregulation function and physical features is determined through exhaustive simulation. Multiple modifications to the original model — delaying thermal transfer, thermoregulation response, etc. — improved accuracy. Across 120+ hours of running, walking, biking and tennis, the method successfully reduces estimation error, leading to collaborations with several companies and researchers in other fields.

We proposed an opportunistic ad-hoc localization algorithm — Urban Pedestrians Localization (UPL) — for estimating the locations of mobile nodes in urban districts. UPL assumes location landmarks are deployed sparsely due to deployment-cost constraints, so most mobile nodes cannot expect to meet landmarks frequently. Each node instead relies on location information received from neighboring nodes to estimate its area of presence. Although the area becomes inexact under motion, it can be used to reduce the areas of presence of others. UPL also leverages information about obstacles such as walls and computes movable areas under mobility for accurate prediction. Experiments showed UPL limits positioning error to 0.7r on average, where r is the radio range.

We also worked on mobile-phone localization for firefighters using WiFi directional antennas placed around a target building, plus several works on relative-position metrics, people-density estimation, and GPS improvement.