Two deep approaches for ADL recognition: A multi-scale LSTM and a CNN-LSTM with a 3D matrix skeleton representation

In this work, we propose a deep learning approach for the detection of the activities of daily living (ADL) in a home environment starting from the skeleton data of an RGB-D camera. In this context, the combination of ad hoc features extraction/selection algorithms with supervised classification approaches has reached an excellent classification performance in the literature. Since the recurrent neural networks (RNNs) can learn temporal dependencies from instances with a periodic pattern, we propose two deep learning architectures based on Long Short-Term Memory (LSTM) networks. The first (MT-LSTM) combines three LSTMs deployed to learn different time-scale dependencies from pre-processed skeleton data. The second (CNN-LSTM) exploits the use of a Convolutional Neural Network (CNN) to automatically extract features by the correlation of the limbs in a skeleton 3D-grid representation. These models are tested on the CAD-60 dataset. Results show that the CNN-LSTM model outperforms the state-of-the-art performance with 95.4% of precision and 94.4% of recall.