Activity recognition using a combination of high gain observer and deep learning computer vision algorithms

•The paper develops a novel method to identify daily living activities of a person using a single wearable sensor on the person's chest.•This paper presents the idea of converting sensor readings to images using spectrograms, so that image processing algorithms can be utilized.•The paper demonstrates how a state estimation observer can highly improve the performance of a deep learning activity recognition algorithm by creating more meaningful input signals for the learning algorithm.•The paper shows how to perform transfer learning from networks pre-trained on millions of images, thus showing how we can train a powerful deep learning network for activity recognition even with just small datasets.•Extensive experimental validation by presenting data from 7 human subjects collected in their home environments.