Search for: All records

The article examines neural network learning of the sparse array configurations in optimum beamforming. Unlike iterative greedy, convex, and global optimization methods for optimum array design, deep learning enables fast reconfigurations of the sparse array in rapid dynamic propagation environments. We employ three different convolutional neural network architectures with varying simplification and parameter counts. The network is trained to select M out of N uniformly spaced antennas to achieve maximum signal-to-interference and noise ratio (SINR) beamforming. Different values of M are considered, including N = 2 M, for studying network performance under an increased number of subarray classes. We consider one desired source and one interference of arbitrary angle, and delineate the learning results for the two cases where the network is trained with the desired source assuming fixed and varying angles. We discuss the benefits of reducing the number of possible configurations due to sidelobe level reductions. It is also shown that the network performance significantly improves with data augmentations and by removing redundant array configurations which produce the same SINR. 

The impact of aspect angle on Doppler effect hinders the capability of a monostatic radar to achieve human activity recognition (HAR) from all aspect angles, i.e., omnidirectional. To alleviate the “angle sensitivity”, sufficient and high-quality training data from multiple aspect angles is mandated. However, it would be time-consuming for the monostatic radar to collect the training data from all aspect angles. To address this issue, this paper proposes a high-quality synthetic data generation algorithm based on high-dimensional model representation (HDMR) for omnidirectional HAR. The aim is to augment a high-quality dataset with collected samples at the radar line-of-sight direction and few samples from other aspect angles. The quality of synthetic samples is evaluated by dynamic time wrapping distance (DTWD) between the synthetic and real samples. Subsequently, the synthetic samples are utilized to train a classifier based on ResNet50 to achieve omnidirectional HAR. Experimental results demonstrate that the averaged HAR accuracy of the proposed algorithm exceeds 91% at different aspect angles. The quality of the synthetic samples generated by the proposed algorithm outperforms two commonly-used algorithms in the literature. 

Human activity recognition (HAR) with radar-based technologies has become a popular research area in the past decade. However, the objective of these studies are often to classify human activity for anyone; thus, models are trained using data spanning as broad a swath of people and mobility profiles as possible. In contrast, applications of HAR and gait analysis to remote health monitoring require characterization of the person-specific qualities of a person’s activities and gait, which greatly depends on age, health and agility. In fact, the speed or agility with which a person moves can be an important health indicator. In this study, we propose a multi-input multi-task deep learning framework to simultaneously learn a person’s activity and agility. In this initial study, we consider three different agility states: slow, nominal, and fast. It is shown that joint learning of agility and activity improves the classification accuracy for both activity and agility recognition tasks. To the best of our knowledge, this study is the first work considering both agility characterization and personalized activity recognition using RF sensing. 

Radar-based recognition of human activities of daily living has been a focus of research for over a decade. Current techniques focus on generalized motion recognition of any person and rely on massive amounts of data to characterize generic human activity. However, human gait is actually a person-specific biometric, correlated with health and agility, which depends on a person’s mobility ethogram. This paper proposes a multi-input multi-task deep learning framework for jointly learning a person’s agility and activity. As a proof of concept, we consider three categories of agility represented by slow, fast and nominal motion articulations and show that joint consideration of agility and activity can lead to improved activity classification accuracy and estimation of agility. To the best of our knowledge, this work represents the first work considering personalized motion recognition and agility characterization using radar.