Transition-aware human activity recognition using an ensemble deep learning framework

Understanding human activities in daily life is of utmost importance, especially in the context of personalized and adaptive ubiquitous learning. Although existing HAR systems perform well-identifying activities based on their inter-spatial and temporal relationships, they lack in identifying the importance of accurately detecting postural transitions that not only enhance the activity recognition rate and reduced the error rate but also provides added motivation to explore and develop hybrid models. It's in this context we propose an ensemble approach of 1D-CNN and LSTM for the task of postural transition recognition, facilitated by wireless computing and wearable sensors. The proliferation of achieving ubiquitous learning will ultimately lead to the creation of adaptive devices enabled by various data analysis and relation learning techniques. Our approach is one of the methods that can be incorporated to enable seamless learning and acquire correlations with adaptive learning techniques. The experimental results on testing datasets including newly produced HAPT (Human Activities and Postural Transitions) show better classification accuracy than existing state-of-the-art HAR approaches (97.84% for transitional activities and 99.04% for dynamic human activities) indicating the capability of the model in ubiquitous learning scenarios and personalized and adaptive human learning environments.