PROJECT TITLE :

A Sensors based Deep Learning Model for Unseen Locomotion Mode Identification using Multiple Semantic Matrices

ABSTRACT:

In recent years, the availability of a variety of sensors within a smartphone has made determining a locomotion mode a task that is not only simple but also extremely convenient. It is helpful to improve journey planning, travel time estimation, and traffic management when more information is available about the mode of locomotion. Although there has been a substantial amount of work done towards the recognition of locomotion modes, the performance of this work is not relevant and heavily depends on the labeled training examples. Because it is impractical to collect previous information (labeled instances) on all possible modes of locomotion, the recognition model should be able to recognize a new or previously unseen mode of locomotion even in the absence of any corresponding training instance. Using labeled training instances, this paper proposes a sensors-based deep learning model that can identify a locomotion mode. In addition to that, the Zero-Shot learning strategy is incorporated into the method in order to determine an unknown locomotion mode. The model generates an attribute matrix by fusing together three different semantic matrices to create the matrix. In addition to this, it will extract the deep learning and hand-crafted features from the training instances in order to construct a feature matrix. In a later step, the model will construct a classifier by first learning a mapping between the attribute matrices and the feature matrices. In conclusion, this research assesses the performance of the method by comparing it to previously collected and published datasets, using accuracy and the F1 score.