Recognition of Activities of Daily Living from Egocentric Videos Using Hands Detected by a Deep Convolutional Network

Ambient assisted living systems aim at supporting older and impaired people using computer technology so that they can remain autonomous while maintaining healthy living. Egocentric cameras have emerged as a powerful source of data to monitor individuals performing activities of daily living since they tend to focus on the area where the current activity takes place while showing manipulated objects and hand positions. While research has focused on activity recognition based on object recognition, this study has investigated the automatic acquisition of additional features modelling interactions between hands and objects using a deep convolutional network. Experiments conducted on a realistic dataset have demonstrated that, not only do those features improve activity recognition, but they can be accurately extracted.