Radar, as one of the sensors for human activity recognition (HAR), has unique characteristics such as privacy protection and contactless sensing. Radar-based HAR has been applied in many fields such as human–computer interaction, smart surveillance and health assessment. Conventional machine learning approaches rely on heuristic hand-crafted feature extraction, and their generalization capability is limited. Additionally, extracting features manually is time–consuming and inefficient. Deep learning acts as a hierarchical approach to learn high-level features automatically and has achieved superior performance for HAR. This paper surveys deep learning based HAR in radar from three aspects: deep learning techniques, radar systems, and deep learning for radar-based HAR. Especially, we elaborate deep learning approaches designed for activity recognition in radar according to the dimension of radar returns (i.e., 1D, 2D and 3D echoes). Due to the difference of echo forms, corresponding deep learning approaches are different to fully exploit motion information. Experimental results have demonstrated the feasibility of applying deep learning for radar-based HAR in 1D, 2D and 3D echoes. Finally, we address some current research considerations and future opportunities.

Automatic modulation classification facilitates many important signal processing applications. Recently, deep learning models have been adopted in modulation recognition, which outperform traditional machine learning techniques based on hand-crafted features. However, automatic modulation classification is still challenging due to the following reasons. Existing deep learning methods are only applicable to the data of the same distribution. In practical scenarios, data distribution is varying with sampling frequency, thus domains with different sampling rates are formed. Besides, it is difficult to construct large-scale well-annotated datasets for all domains of interest. We define the domain with sufficient data as the source domain, while the domain with insufficient data as the target domain. Obviously, the classification model performs weakly in the target domain. To address these challenges, we propose an adversarial transfer learning architecture (ATLA), incorporating adversarial training and knowledge transfer in a unified way. Adversarial training performs an asymmetric mapping between domains and reduces the domain shift. Knowledge transfer is used to mine prior knowledge from the source domain. Experimental results demonstrate that the proposed ATLA substantially boosts the performance of the target model, which outperforms the existing parameter-transfer approach. With half of the training data reduced, the target model achieves competitive recognition accuracy to supervised learning. With one-tenth of training data, the promoted accuracy is up to 17.3% points.

Accurate beam tracking in high-mobility vehicular networks has become a key issue for reliable communication link establishment. Recent work on integrated sensing and communication has shown the superiority of the sensing-assisted beam tracking. By estimating the angular parameters of vehicles based on the radar echoes, the roadside unit (RSU) can perform a so-called predictive beamforming to pair the beams with high accuracy as well as low latency. To further reduce the estimation error due to the nonlinear measurements, in this letter, we develop a deep learning-based approach to exploit the dependency between the angles and the measurements, and to obtain the angle estimates. Simulation results show that our proposed approach can improve the estimation performance and maintain reliable communications in high-mobility vehicular networks.

Currently, developments in wireless sensing technologies have shown that wireless signals can be employed to transmit information between wireless communication devices and are also able to realize passive target wireless sensing. Wireless sensing has diverse Internet-of-Things applications in indoor human detection, such as in device-free localization, activity recognition and fall detection, respiration detection, gait recognition, user identification, and so forth. Deep learning (DL), with the latest breakthroughs in machine learning (ML) and artificial intelligence (AI), seems to be a feasible technique for device-free wireless sensing (DFWS) and human detection in a more intelligent and autonomous manner. Although DL has attracted wide spread attention in computer vision (CV), AI games, speech recognition, automated vehicles, and other fields, its application in wireless sensing systems (WSSs) is relatively new, and little attention has been paid to it. Motivated by these developments, this article clarifies the motivation and mechanism of the DL-aided WSSs for human detection. First, we survey the most advanced architecture of DL that may be powerful for WSSs. We also review conventional ML and DL approaches to human detection based on red green blue (RGB)/depth camera and radar: one reason is to introduce the successful experience in these areas to the field of wireless sensing and another reason is that the possibility of combining and fusing information from the heterogeneous types of sensors is expected to improve the overall performance of practical human detection systems. We provide a comprehensive survey of the state-of-the-art research on wireless sensing for human detection with a focus on WSSs. Furthermore, a general structure of the DL-based WSS is introduced in detail for hitherto unexplored applications and future wireless sensing scenarios. We also discuss some open research issues in wireless sensing for human detection, including data acquisition for DL model training, calibration of signals from commercial devices, multimodal sensing, simultaneous user identification and activity recognition, multiuser human detection, and generalization ability of DL models, to indicate future research directions.

Gesture recognition is the central enabler of human-computer interaction (HCI). In addition to the semantic information contained in gestures, gesture-based user identification can effortlessly enhance HCI system security. Recently, the Wi-Fi-integrated sensing and communication (ISAC) technology has shown great potential in a field hitherto occupied by computer vision and radar sensing. In this work, leveraging Wi-Fi sensing, we propose a system called WiGesID that achieves joint gesture recognition and human identification (JGRHI). The basic idea behind WiGesID is to identify personalized spatiotemporal dynamic patterns from the gestures of different users. Moreover, we develop an effective approach to recognize new categories of gestures and users by computing relation scores between the features of the new category samples and the support samples. To evaluate the performance, we implemented WiGesID and conducted extensive experiments. The results demonstrate that our system outperforms the state-of-the-art method for cross-domain sensing and accurately recognizes new categories, which promotes the use of this application of Wi-Fi sensing in HCI.

The downlink communications are vulnerable to intelligent unmanned aerial vehicle (UAV) jamming attack. In this paper, we propose a novel anti-intelligent UAV jamming strategy, in which the ground users can learn the optimal trajectory to elude such jamming. The problem is formulated as a stackelberg dynamic game, where the UAV jammer acts as a leader and the ground users act as followers. First, as the UAV jammer is only aware of the incomplete channel state information (CSI) of the ground users, for the first attempt, we model such leader sub-game as a partially observable Markov decision process (POMDP). Then, we obtain the optimal jamming trajectory via the developed deep recurrent Q-networks (DRQN) in the three-dimension space. Next, for the followers sub-game, we use the Markov decision process (MDP) to model it. Then we obtain the optimal communication trajectory via the developed deep Q-networks (DQN) in the two-dimension space. We prove the existence of the stackelberg equilibrium and derive the closed-form expression for the stackelberg equilibrium in a special case. Moreover, some insightful remarks are obtained and the time complexity of the proposed defense strategy is analyzed. The simulations show that the proposed defense strategy outperforms the benchmark strategies.