Zhaocheng Yang

An indoor multi-target tracking method by exploiting target motion characteristics and prior boundary knowledge based on multiple-input multiple-output (MIMO) radar is presented in this paper. The main task is to consistently and accurately track the people in the room. The challenge of indoor multi-target tracking lies in the mutual interference among target echo signals and the influence of multipath effects. In this paper, we first conduct long-short term preprocessing to remove static clutter, separate strength and weakness targets, and generate strength and weakness measurements. Meanwhile, we utilize prior boundary knowledge to eliminate multipath effects outside the boundary. Then, we manage trajectories based on the motion characteristics of targets and adopt different association and tracking strategies according to the trajectory types. Experimental results demonstrate that the proposed method achieves tracking errors of less than 0.1 m in various scenarios, with an average accuracy of 97.26% in people counting, which achieves not only high-precision target tracking but also superior performance in people counting compared to existing methods.

The existing radar-based people counting methods have achieved comparatively good detection performance in most scenarios. But in the doorway, opening and closing the door creates interfering trajectories, which leads to the confusion between person and door, and reduces the people counting performance. To address this problem, we propose a bi-direction people counting method exploiting door open-close discrimination using 4D millimeter wave radar. We first detect the presence of door open-close events within the region of interest based on target detection and point cloud height feature extraction. For point clouds captured during door movement, we construct a rectangular frame to separate door target point clouds and perform clustering analysis. Then, we explicitly incorporate door targets into the multi-target tracking process. To achieve more accurate tracking, the data association is decomposed into three stages: type matching, cluster matching, and point assignment within clusters. Besides, different trajectory management strategies are applied for human targets and door targets. Finally, we count people according to the trajectory information and edge region. The experimental results show that the proposed method achieves an average detection accuracy for door open-close events of over 99%, with a counting accuracy of 97.43% for people entering and exiting in passageways. Moreover, it can maintain a counting accuracy of 96% in scenarios with door interference, demonstrating good counting robustness.

Conventional radar-based sit-to-stand (STS) estimation methods typically require subjects to perform the movement at specific locations or follow predefined directions, which limits their applicability in real home environments and significantly restricts the natural freedom of user behaviors. Moreover, most existing approaches treat the human body as a whole, neglecting the micro-Doppler characteristics of local limbs. In this paper, we explore a millimeter-wave radar-based STS estimation and person identification method that allows users to perform the STS action casually in home scenarios, while dynamically capturing motion characteristics and extracting critical local limb micro-Doppler signatures for person identification. The core idea is to segment limb regions from the point cloud data, identify motion directions that exhibit high-quality micro-Doppler features, and estimate radial velocities of the corresponding local limbs. Finally, by utilizing the radial velocity of limbs in combination with Doppler features, person identification is performed using long short-term memory (LSTM) networks. Experimental results demonstrate that the proposed method effectively extracts limb-level radial velocities, with most intraclass correlation coefficients (ICCs) exceeding 0.9. Furthermore, by applying the extracted velocity features to person identification for eight individuals, we achieve an accuracy of 95.3%, highlighting its potential applications in home monitoring systems.

Indoor radar human sensing has received great attention because of its robustness to darkness, sensitivity to human motion, and no privacy issues. However, its performance is susceptible to interference from multipath and shadow ghosts and remains a challenge. In this paper, we propose a space boundary-aware indoor people tracking and counting method using multi-input multi-output (MIMO) radar. We first perform data screening from a large number of radar observations to screen the 0-person data and 1-person data. Then, we utilize the antenna beam characteristics to estimate the radar front space boundary using the 0-person data and the multipath reflection characteristics to estimate the radar side space boundary using the 1-person data. The obtained space boundaries are applied to the people tracking process to eliminate multipath and shadow ghosts. The proposed method can achieve front boundary estimation accuracy with a mean absolute error (MAE) of 0.084m, side boundary estimation accuracy with an MAE of 0.184m, people counting accuracy of 85.63%, and multi-person tracking MAE of 0.071m.

Radar sleep posture recognition has received much attention due to its advantages such as privacy preservation and freedom from quilt occlusion. However, radar echo signals are highly sensitive to different humans, which leads to poor generalization ability. To address this problem, we propose a human sleep posture recognition method exploiting spatial morphology and local micro-Doppler characteristics using multiple-input multiple-output (MIMO) radar. We first perform multidimensional imaging of human targets. Then, we extract the spatial morphology features of human sleep posture and local micro-Doppler features of human sleep life activities and utilize the Bagged Trees classifier to complete the classification of left lateral, right lateral, and non-lateral. For the non-lateral posture, we further extract new feature sequences from multidimensional imaging and combine them with a long and short-term memory (LSTM) network to complete the classification of supine and prone posture. Experimental results show that the proposed method achieves an average accuracy of 97.6% for recognizing the four sleep postures, has good reliability across individuals, and performs efficiently in most types of beds with single human at different sleeping positions. Finally, we have deployed the proposed method in a designed real-time system, validating its low computation complexity and the feasibility of edge applications.

In 5G new radio (NR) integrated sensing and communication (ISAC) systems, the overhead of sensing time resources in waveform design is a key metric. Traditional uniform pulse sensing waveforms, which require fixed time slot positions, incur large sensing overhead and may not be feasible within the context of existing 5G NR standards for low-latency scenarios. To address this issue, two novel frame-level sensing waveform designs based on double coprime sampling structures are proposed via downlink sensing. Specifically, two types of novel orthogonal frequency-division multiplexing (OFDM) structure pulse sensing waveforms based on coprime sampling are first proposed to accommodate the specific 5G NR frame structure. Then, a low-complexity frequency-domain radar signal processing method is proposed for sensing target detection and velocity estimation based on the constructed equivalent virtual fast-slow-time domain signal. Moreover, the theoretical sensing performance of the proposed waveforms and the computational complexity of the sensing method are thoroughly analyzed. Compared with the conventional uniform OFDM-based waveform design, the proposed coprime-based frame-level OFDM sensing waveforms utilize significantly fewer time slots, resulting in reduced sensing overhead with slight sensing performance loss. Numerical examples are presented to illustrate that the proposed waveforms can offer robust sensing performance with reduced sensing overhead and have a computational complexity a little higher than that of conventional fast Fourier transform algorithms.