In recent years, significant research has been conducted on video-based human pose estimation (HPE). While monocular two-dimensional (2D) HPE has been shown to achieve high performance, monocular three-dimensional (3D) HPE poses a more challenging problem. However, since human motion happens in a 3D space, 3D HPE offers a more accurate representation of the human, granting increased usability for complex tasks like analysis of physical exercise. We propose a method based on MediaPipe Pose, 2D HPE on stereo cameras and a fusion algorithm without prior stereo calibration to reconstruct 3D poses, combining the advantages of high accuracy in 2D HPE with the increased usability of 3D coordinates. We evaluate this method on a self-recorded database focused on physical exercise to research what accuracy can be achieved and whether this accuracy is sufficient to recognize errors in exercise performance. We find that our method achieves significantly improved performance compared to monocular 3D HPE (median RMSE of $30.1$ compared to $56.3$, p-value below ${10}^{-6}$) and can show that the performance is sufficient for error recognition.

Developments in radio detection and ranging (radar) technology have made hand gesture recognition feasible. In heat map-based gesture recognition, feature images have a large size and require complex neural networks to extract information. Machine learning methods typically require large amounts of data and collecting hand gestures with radar is time- and energy-consuming. Therefore, a low computational complexity algorithm for hand gesture recognition based on a frequency-modulated continuous-wave (FMCW) radar and a synthetic hand gesture feature generator are proposed. In the low computational complexity algorithm, two-dimensional Fast Fourier Transform is implemented on the radar raw data to generate a range-Doppler matrix. After that, background modelling is applied to separate the dynamic object and the static background. Then a bin with the highest magnitude in the range-Doppler matrix is selected to locate the target and obtain its range and velocity. The bins at this location along the dimension of the antenna can be utilised to calculate the angle of the target using Fourier beam steering. In the synthetic generator, the Blender software is used to generate different hand gestures and trajectories and then the range, velocity and angle of targets are extracted directly from the trajectory. The experimental results demonstrate that the average recognition accuracy of the model on the test set can reach 89.13% when the synthetic data are used as the training set and the real data are used as the test set. This indicates that the generation of synthetic data can make a meaningful contribution in the pre-training phase.

The application of radar in the medical field has attracted a lot of attention recently. Vital sign detection and monitoring are essential for certain use cases. Due to the limitations and discomfort during contact-based vital sign monitoring, the non-contact approach becomes preferred. In this work, a signal processing approach for multi-target breathing rate and heart rate monitoring is proposed. The proposed approach uses a frequency modulated continuous wave (FMCW) radar for contact-less vital sign monitoring. Using the proposed approach, the vital signs of multiple persons can be monitored simultaneously using a single mmWave radar. The results obtained using the proposed approach show an excellent match compared to a medical-grade instrument. This is especially notable for the heart rate monitoring, where the achieved accuracy error is only 7.77 %, outperforming other similar works.

The ImageNet moment was a turning point for Convolutional Neural Networks (CNNs), as it demonstrated their potential to revolutionize computer vision tasks. This triumph of CNNs has motivated solving even more complex problems involving multiple tasks from multiple data modalities. Conventionally, a single CNN accelerator has been optimized to perform just one task or multiple correlated tasks. This study presents a shared-layers approach that leverages the pattern-learning capabilities of CNNs to perform multiple uncorrelated tasks from different modalities using a single hardware accelerator. We overcame the challenge of data imbalance in multi-modal learning by synthetic data generation. We achieved an average classification accuracy above 90% on a single CNN accelerator, which would otherwise require three accelerators. Due to the reliability concerns imposed by transistor shrinking and aging, we extended the shared layers methodology and introduced a fault-tolerant CNN accelerator with reconfigurable capabilities supporting fault-tolerant (FT), high-performance (HP), and de-stress (DS) modes. FT mode provides high reliability against soft errors utilizing double/triple modular redundancy, HP mode offers peak performance of 0.979 TOPs using parallel execution, and DS mode reduces dynamic power consumption by up to 68.6% in clock-gated design and even more using a partial reconfiguration method, contributing to decelerating the aging process of the circuit. We have comprehensively evaluated two different CNN architectures (i.e., fused and branched), for three distinct tasks, in three different operating modes, based on accuracy, quantization, pruning, hardware resource utilization, power, energy, performance, and reliability.