POIDEN: position and orientation independent deep ensemble network for the classification of locomotion and transportation modes

Sensor-based recognition of locomotion and transportation modes has numerous application domains including urban traffic monitoring, transportation planning, and healthcare. However, the use of a smartphone in a fixed position and orientation in previous research works limited the user behavior a lot. Besides, the performance of naive methods for position-independent cases was not up to the mark. In this research, we have designed a position and orientation independent deep ensemble network (POIDEN) to classify eight modes of locomotion and transportation activities. The proposed POIDEN architecture is constructed of a Recurrent Neural Network (RNN) with LSTM that is assigned the task of selecting optimum general classifiers (random forest, decision tree, gradient boosting, etc.) to classify the activity labels. We have trained the RNN architecture using an intermediate feature set (IFS), whereas, the general classifiers have been trained using a statistical classifier feature set (SCFS). The choice of a classifier by RNN is dependent upon the highest probability of those classifiers to recognize particular activity samples. We have also utilized the rotation of acceleration and magnetometer values from phone coordinate to earth coordinate, proposed jerk feature, and position insensitive features along with parameter adjustment to make the POIDEN architecture position and orientation independent. Our team "Gradient Descent" has presented this work for the "Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge".