Motion Estimation of Plush Toys Through Detachable Acceleration Sensor Module and Machine Learning

We propose a system that estimates motion in a plush toy by means of an attached sensor device and gives the user a sound feedback corresponding to the predicted motion. We have created several different types of detachable acceleration sensor modules as an accessory for the toy. This module can be attached at any position on a commercially available plush toy. The user can create original motions by teaching through demonstration, and the captured sensor data is converted into 2D image data. We extracted the histograms of oriented gradients (HOG) features and performed learning with a support vector machine (SVM). In an evaluation, we decided the attaching parts and motions in advance, and participants moved a plush toy in accordance with these. Results showed that it was possible to estimate the plush toy’s motion with high accuracy, and the system was able to register a sound for each motion.