Sparse Feature Learning for Human Activity Recognition
Shan Ullah, Deok‐Hwan Kim · 2021
In this paper, we propose an end-to-end deep learning model for human activity recognition. Our model is equipped with sparse learning, which absorbs a greater number of classes without making a significant change in the size of the model while sustaining the accuracy of existing classes. In addition, our model is lightweight than state-of-the-art models as we have utilized FCN-LSTM (Fully convolution network - Long Short-term Memory). Our model predicts human activities such as walking, walking-upstairs, walking-downstairs, sitting, standing, and laying (total 6 classes). For validation of our deep learning model, we have utilized a well-known opensource dataset such as the UCIHAR-dataset, which contains collections of smart-phones data of 30-subjects performing different activities with a smartphone. We evaluated the model using sparse learning and have shown that our model outperforms in learning features with few epochs with high accuracy and compact size, and efficient inference time correspondingly.