Efficient dense labelling of human activity sequences from wearables using fully convolutional networks

Highlights：

• A new method to address the multi-class windows problem in human activity recognition from sequences of activity data.

• Propose a fully convolutional network architecture for dense labelling and prediction of sequences of arbitrary length.

• The convolutional network method is much more efficient than CNN counter-parts.

• Release of a new activity dataset collected from hospitalised older people.

• Demonstrate the generalisability of the method on three datasets using sample- and activity-based measures.