About recognition of the movements made with use of the mobile device

December 5, 2017

          Our engineers made experiments to determine usage of the mobile device by using positioning sensors embedded in mobile device.

          For an experiment data sets were generated from accelerometer and magnetometer measurements collected on Android device. Collected data represented record of indications of a sensor in case of execution of the given set of movements of a hand holding mobile device. At the following stage different approaches to data representation were evaluated  with neural MPLClassifier network of Scikit-Learn library.

          The most demonstrative are the results received when using data representations of a sensor as a set of the discrete values belonging to a certain class (movement) and as the indivisible vector representing the sequence of the measurements written in runtime of movement. These approaches yield the positive results within 70-99% of cases at different volume and quality of learning selection.

          Further, our company will continue collection of learning data and the thin setup of the classifier to increase robustness and accuracy that stably exceeds 90%. It is in addition planned to execute assessment of quality of determination of actions upon transition from the rotation matrix received on getRotationMatrix sensor output to quaternions.