Ensemble of deep autoencoder classifiers for activity recognition based on sensor modalities in smart homes

Over the past few years, a particular interest has been focused toward activity recognition domain. Indeed, human activity recognition pays more attention on the extraction of relevant and discriminative features whose the implementation facilitates the seamless monitoring of functional inhabitant abilities with the involvement of sensing technology in the smart home environments. However, despite the exponential efforts made by individual standard machine learning techniques, and recently by the remarkable breakthrough of deep learning methods, designing robust activity recognition architecture remains a major challenge in term of performance, due to a high degree of uncertainty and complexity caused by inherent behavior of human actions. For the former case, the drawbacks are essentially composed of heuristic and hand-crafted methods for features extraction, shallow features learning, and learning of low amount of well-labeled data. While the latter suffers from imbalanced datasets and problematic data quality in real-life datasets. In addition, the choice of suitable sensor types is also critical for successful human activity recognition. This paper proposes an ensemble of deep classifier techniques based on hybrid sensor types composed of wearable and environment interactive sensors to improve the prediction and recognition performance of activities of daily living in smart home environments. Indeed, this ensemble is designed by combining the both automatically learned features and hand-crafted features from Denoising Stacked Autoencoders (DSAE) and Random Forest (RF) algorithm respectively. Specifically, the combination involves both the features and outputs of the two techniques using stacking learning. The use of two public benchmark datasets has enabled to evaluate our approach. Furthermore, the experimental results show the accuracy improvement of the ensembles of deep autoencoders classifiers compared to denoising stacked autoencoder networks and random forest algorithm performed individually. Hence, our approach adaptability to ubiquitous environments and its effectiveness in the recognition of human activity applications.