Enhancing Gait Recognition System Using Smartphone Sensors (IMU) and Deep Learning Techniques

Abstract

In this article, person re-identification approach was developed based on structured data that was obtained from smartphone sensors, especially gyroscopes and accelerometers. The approach basically uses sensors measurements as gait features to identify the person. The data is structured in tables but converted to images with specific dimensions (n*m) where  is number of features (6: 3 gyroscopes components and 3 accelerometers components) and  the sample length; it refers to period of gait motion. The dataset was from OU-ISIR Biometric Database. The parameter  was chosen after many tries and best value was 128. The structured data is converted to images with dimensions (128*6) and then feed into developed deep learning model that basically uses deep learning layers (Convolution, Batch normalization, Memory units, Masking, Dropout, Permute and Pooling). The tuning process was processed on some parameters to get best values such as number of units in LSTM (Long Short-Term Memory), strides of convolution layers in the model and learning rate of learning process. In addition, parallel processing in the model was used to enhance the data features since the number of features is small, the final model showed high re-identification accuracy (98.23 %)