Data filtering and deep learning for enhanced human activity recognition from UWB radars

Human activity recognition (HAR) is among the most popular research topics. Indeed, recognizing human activities can help provide appropriate assistance to older adults and address the challenges of an aging population. Hence, HAR solutions based on Ambient Intelligence (AmI) have been proposed to face the challenges of an aging population. In addition, we denoted an increasing interest in Ultra-WideBand (UWB) radars for HAR. In this work, we exploited three UWB radars to recognize 15 human activities performed by 9 participants in a prototype apartment. One of the main contributions of this paper is the improvement of classification results compared to our previous work, with an average accuracy of 26% higher for the top-1 classification. To do so, we emphasize the data cleaning stage. More precisely, since the amount of data is insignificant, data provided by UWB radars have been cleaned with a well-known band-pass Chebyshev type I filter of order 2. Applying that kind of filter to data provided by UWB radars is uncommon in the literature. In addition, two different deep learning architectures to classify the cleaned data have been exploited. The first is a relatively simple Convolutional Neural Network (CNN), and the second is Efficient-CapsNet. We obtained similar performances between these two architectures for the top-1 with an accuracy of approximately 64%. However, from the top-2 to top-5, the CNN outperformed Efficient-CapsNet.