GraphConvLSTM: Spatiotemporal Learning for Activity Recognition with Wearable Sensors

Wearable activity recognition is an important area for healthcare applications, especially for children and elderly people. Current research proves that deep learning techniques are capable of learning deep features from the sensor raw signal. However, this task faces two challenges. First, it is hard to capture the dynamic and non-linear interactions between different sensor modalities and time slots. Besides, the spatial relationship between sensors convey significant information for wearable activity recognition, thus how to model the sensor relationship is one of the problems that need to be solved. In this work, we propose GraphConvLSTM, a scalable framework that model spatial relationship, local interaction, and long-term temporal dependency simultaneously. We evaluate our model on two real-world activity recognition datasets, and it achieves 1.75%, 2.68% accuracy improvements over existing methods (accuracy: 90.83%, 93.08%) for wearable activity recognition task, which demonstrates the ability of the proposed approach to learn spatial, local and global temporal features.