Deep Learning on Sensor-based Activity Recognition

Tsuyoshi Okita, Sozo Inoue,
GPU Technology Conference
(Not Available)
(Not Available)
This poster shows our deep learning efforts on activity
recognition which handles smartphone and environmental sensors. We
aim at making an analysis on the IoT data which are human
understandable as far as possible. We have three
engines: (1) the compositional Convolutional Neural Network+Long
Short-Term Memory model which can handle multiple objects (Ubicomp
2017 poster), (2) the ensemble Recurrent Neural Network which can
calculate variable importance (UBI kenkyukai 56, 2017), and (3) random
forests. We do the model selection / Bayesian optimization on these
three Machine Learning methods to obtain the better results. We also
make a trial to obtain the visualization of the sensor inputs where we
do an analysis with the pose estimation technique using additional visual data
which we additionally obtained during the process. This poster is a summary of two presentations this year
(Ubicomp 2017 poster; UBI kenkyukai 56) and one on-going work.

Data Files

No records to display