Deep learning based human age estimation using inertial data of normal walk

Abstract

Human gait enables information about soft biometrics features in multiple scenarios where the environment and visual cues are not known. This characteristic of human gait has numerous application in the field of surveillance, advertisement and monitoring and provides a practical approach in numerous real world applications such as motion sensor, alarms and detections. Moreover, since human gait style varies from person to person, the human gait is believed to be unique. To estimate a person’s age by using non-visual features is a very effective task and can help us in various domains like, soft-biometrics based applications, forensic medicine and forensic anthropology, human computer interaction (HCI), surveillance (where visual cues are hidden), criminal investigation, targeted campaigns and advertisement, security and safety management systems etc. In this thesis, we propose a deep learning based model to analyze inertial human gait data in order to estimate subject’s age. We used our previous gait data set collected for 40 subjects using on-body inertial sensors. Each step and stride are collected for 40 subjects (26 Males:14 Females) with accelerometer and gyro meter reading in the 3d (x,y,z) planes. We employ Long Short Term Memory (LSTM) networks to utilize the memory of the networks for adjacent steps and strides of a subject to estimate subject’s age. The results are generated by raw data vectors and normalized vectors for the LSTM. We observe that normalized data is able to converge better with small batch size. The classification results in 87% accuracy with different training windows and batch sizes