Gait verification using deep learning models, accelerometers and gyroscope data

Sammanfattning: Biometrical authentication systems, with its advantages over the traditional methods such as tokens and cards combined with secret passwords and pins, are becoming popular these days. In addition recent advances in machine learning helped improve the performance of biometrical systems, that can now be used commercially. The scope of the present thesis work is to investigate what is the minimum sample length for users’ authentication based on monitoring of walking manner (gait). In the context of the thesis three deep learning models of different architectures were built and trained using 6 dimensional inertialsensor data, to assess the accuracy and the reliability of the proposed methods. The models were trained using 1, 2, 4 and 8 step gait patterns. The results showed that the most accurate model was the CNN+LSTM for short gaits, while for longer gaits the performance of the models seemed to converge. At the same time all the models performed better for longer gaits, with relatively small differences.