Currently, Human Activity Recognition (HAR) applications need a large volume of data to be able to generalize to new users and environments. However, the availability of labeled data is usually limited and the process of recording new data is costly and time-consuming. Synthetically increasing datasets using Generative Adversarial Networks (GANs) has been proposed, outperforming cropping, time-warping, and jittering techniques on raw signals. Incorporating GAN-generated synthetic data into datasets has been demonstrated to improve the accuracy of trained models. Regardless, currently, there is no optimal GAN architecture to generate accelerometry signals, neither a proper evaluation methodology to assess signal quality or accuracy using synthetic data. This work is the first to propose conditional Wasserstein Generative Adversarial Networks (cWGANs) to generate synthetic HAR accelerometry signals. Furthermore, we calculate quality metrics from the literature and study the impact of synthetic data on a large HAR dataset involving 395 users. Results show that i) cWGAN outperforms original Conditional Generative Adversarial Networks (cGANs), being 1D convolutional layers appropriate for generating accelerometry signals, ii) the performance improvement incorporating synthetic data is more significant as the dataset size is smaller, and iii) the quantity of synthetic data required is inversely proportional to the quantity of real data.

Activity Recognition (AR) is an active research topic focused on detecting human actions and behaviours in smart environments. In this work, we present the on-line activity recognition platform DOLARS (Distributed On-line Activity Recognition System) where data from heterogeneous sensors are evaluated in real time, including binary, wearable and location sensors. Different descriptors and metrics from the heterogeneous sensor data are integrated in a common feature vector whose extraction is developed by a sliding window approach under real-time conditions. DOLARS provides a distributed architecture where: (i) stages for processing data in AR are deployed in distributed nodes, (ii) temporal cache modules compute metrics which aggregate sensor data for computing feature vectors in an efficient way; (iii) publish-subscribe models are integrated both to spread data from sensors and orchestrate the nodes (communication and replication) for computing AR and (iv) machine learning algorithms are used to classify and recognize the activities. A successful case study of daily activities recognition developed in the Smart Lab of The University of Almería (UAL) is presented in this paper. Results present an encouraging performance in recognition of sequences of activities and show the need for distributed architectures to achieve real time recognition.

Nowadays, Smart Environments (SEs) enable the monitoring of people with physical disabilities by incorporating activity recognition. Thermal cameras are being incorporated as they preserve privacy. Some Deep Learning (DL) solutions use the pose of the users because it removes external noise. Although there are robust DL solutions in the visible spectrum, they fail in the thermal domain. Thus, we propose THPoseLite (Thermal Human Pose Lite), a Convolutional Neural Network (CNN) based on MobileNetV2 that extracts pose from Thermal Images (TIs). In a novel way, an auto-labeling approach has been developed. It includes a background removal using an optical flow estimator. It also integrates Blazepose (a pose estimator for Visible spectrum images (VSIs)) to obtain the poses in the pre-processed TIs. Results show that the pre-processing increases the percentage of detected poses by Blazepose from 19.55% to 76.85%. This allows the recording of Human Pose Estimation (HPE) datasets in the visible spectrum without requiring visible spectrum cameras or manually annotating datasets. Furthermore, THPoseLite has been embedded in an Internet of Things (IoT) device incorporating an edge Tensor Processing Unit (TPU) accelerator, which can process TIs recorded at 9 Frames Per Second (FPS) in real-time (12.28 FPS). It requires fewer than 6W of energy to run. It has been achieved using model quantization, decreasing the accuracy in estimating the poses by only 1%. The MSE of MobileNetV2 in test images is 35.48, obtaining accurate poses in 21% of the images that Blazepose is not able to detect any pose.