Towards Improved Drink Volume Estimation Using Filter-Based Feature Selection

Abstract

Maintaining adequate hydration is crucial for ensuring positive health outcomes. A variety of solutions have been proposed to assist individuals in achieving optimal water intake, including smart-bottles with embedded consumption tracking. To improve user convenience, we have proposed an attachable solution aimed at providing retrofittable smart-bottle functionality. This paper summarizes recent progress towards improving the volume estimation accuracy of this attachable device. Namely, four filter-based feature selections tools are employed to rank a superset of attributes describing the inclination trajectory of the bottle as estimated from an accelerometer sensor. By sequentially constructing feature sets of varying order based upon the rankings provided by each algorithm, binary volume classification accuracy is increased versus our previously employed feature set for each of the considered algorithms. Results are demonstrated using a newly-constructed set of 1,200 unique drink events for partitions of varying volume separations. For the median partition case, error rate is decreased for the best-case set by 5%. For a partition along the upper and lower quartiles, a best-case improvement of over 10% is achieved.