A Flexible Approach for Human Activity Recognition Based on Broad Learning System

Deep Learning (DL) based methods have recently been receiving attention in Human Activity Recognition (HAR) for their strong capability of nonlinear mapping. However, these methods suffer from high time consumption during training process due to enormous network parameters. Moreover, the DL-based scheme is less capable of incremental learning which is important for some online human activity recognition applications. In this paper, the Broad Learning System (BLS) known as a promising alternative to DL-based methods is introduced to the classification of human activities. Both the online and offline BLS-based recognition frameworks are proposed to enhance the system flexibility. Specifically, during the online training stage, the artificial hyperspherical data generation model is incorporated into the incremental BLS, enabling it to update the model to accommodate new incoming data more efficiently. Experiments are made towards the proposed BLS network based upon two public human activity datasets, namely, HART and WISDM. The results demonstrate the advantage of the proposed BLS-based scheme over the classic DL-based approaches in terms of the training speed and prediction accuracy.