Towards a Multi-Sensor Paradigm for High Resolution Infant Activity Classification

This paper proposes a multi-sensor framework for high resolution classification of infant activities. This involves identifying various activities that occur before onset of walking, including crawling, cruising, sitting, falling, and standing. This is achieved using a system involving multiple accelerometer sensors placed on the infant's body. The sensors used in this setup are lightweight to ensure minimum discomfort. In addition, the sensors used are low-cost and consume very low power that makes the system suitable for practical deployments. The collected data is preprocessed and fed to an Artificial Neural Network module to classify the different activities. A hierarchical binary classification strategy is used in order to ensure fine-grain classification of the infant activities. The proposed setup is experimentally validated using real-time data from two participants collected over multiple days. A characterization of the system is provided to understand the trade-offs between system complexity and classification performance. Furthermore, with extensive experimentation, an analysis of the system performance with respect to the location of the sensor placed on the infant's body is provided.