In this letter, we introduce ultrawideband (UWB) Pose, a through-wall human pose reconstruction framework using UWB multiple-input--multiple-output (MIMO) radar. Previous radio frequency-based works have achieved two-dimensional (2-D) and 3-D human pose reconstruction in human detection of occlusion scenes. However, these methods are difficult to adapt to the environment and often suffer from high computational complexity. To address this issue, we first construct 3-D radar images of human targets using UWB MIMO radar system, and transform those radar images into discrete 3-D point data. Then, design a lightweight deep learning network to extract human body features from the input point data, finally convert the features into 3-D pose coordinates. The comparative experimental results show that the human pose reconstruction error of our UWB-Pose framework can be as low as 38.84 mm. Importantly, the number of parameters and floating point operations are further reduced to 1.02 M and 2.75 G, to meet the needs in practical deployments and the scope of applicability.

Radar-based human activity sensing possesses the advantage of penetrating detection, making it highly promising for applications in the fields of security, rescue, and medical treatment. With the advancement of deep learning technology, there is a growing interest in the field of human activity sensing with radars, including tasks such as action recognition and pose reconstruction. However, existing research usually treats these two problems as separate tasks. In this work, we propose a dual-task framework for jointly reconstructing human three-dimensional (3D) pose and classifying human action from 4D radar images. Our proposed framework utilizes an ultra-wideband multi-input multi-output (MIMO) radar as the detection sensor to obtain the range-azimuth-height-time 4D imaging data of human targets. A human pose reconstruction network based on 3D convolutional neural network (CNN) is then employed to reconstruct the 3D human pose, while a dual-branch network based on multi-frame 3D human poses and 4D radar image is used to classify the human action. To evaluate the performance of the proposed framework, a dual-task dataset is constructed by merging 4D imaging radar and camera data. Experiments and multi-scenario measurements are constructed to validate the effectiveness of the proposed dual-task framework. The results demonstrate that the proposed dual-task network significantly improves the accuracy of both tasks, while providing a single solution for human action recognition and pose reconstruction.

Radar plays an increasingly important role in target detection because of its all-weather detection capability. Furthermore, to detect weak and flexible target, effectively coherent integration method is required in radar signal processing to enhance the target signal intensity. Based on the Radon-Fourier transform (RFT) algorithm, which is widely used to solve the problem of weak target detection, range migration and Doppler ambiguity in low signal-to-noise ratio (SNR) environment, an improved multiple pulse repetition intervals (PRIs) RFT (MRFT) coherent integration method is presented in this paper. The method adopts the Linear frequency modulated (LFM) signal with three different PRIs as the transmitted signal, then combines with the RFT outputs of three RPIs to enhance the target energy and suppress the blind speed side lobe (BSSL) caused by RFT. Meanwhile, the design rules and constraints of this method are also discussed in detail. The experiments with simulation data and real measured data are provided to demonstrate that the designed method can get better BSSL suppression capability and higher detection rate than conventional methods.

Human action recognition (HAR) methods based on ultra-wideband (UWB) multiple-input multiple-output (MIMO) radar have demonstrated substantial potential in complex environments. However, the use of single-modal data, such as spectrograms or point clouds, poses significant limitations; it is easy to get confused when dealing with similar actions, particularly in challenging scenarios. To overcome these limitations, this study proposes an action recognition approach that leverages bimodal data and spatiotemporal features for improved performance. Four-dimensional radar image data of the human body are captured using UWB MIMO radar and processed into four-dimensional point cloud data and four-dimensional intensity projection image data, which collectively serve as bimodal feature inputs. To process these bimodal feature data, a network architecture named Point Image Fusion-Net (PIF-Net) is developed, incorporating spatiotemporal sequence classification to achieve recognition of human actions. The network initially extracts spatial features from the bimodal feature data, followed by temporal feature extraction, culminating in the classification of spatiotemporal features. Experimental results show that the proposed method achieves an accuracy of 96.75% in action recognition tasks, validating the approach.

Accurate crowd flow prediction is essential for traffic guidance and traffic control. However, the high nonlinearity, temporal complexity, and spatial complexity that crowd flow data have makes this problem challenging. This research proposes a dynamic graph convolutional network model (Res-DGCN) based on the residual network structure for crowd inflow and outflow prediction in urban areas. Firstly, as the attention layer, the spatio-temporal attention module (SA) is employed for capturing the spatial relationship between the target node and the multi-order adjacent nodes by processing the features of the human flow data. Secondly, a conditional convolution module (SCondConv) is used to enhance the model’s capacity for learning about the shifting characteristics of crowd flow to obtain spatial dependence. Finally, we train the model with the Huber loss function to lower the model’s sensitivity to outliers and achieve optimal convergence. In two public datasets, the mean absolute error (MAE) of the proposed model is improved by 5.2% and 9.4%, respectively, compared to the baseline models, and the root mean square error (RMSE) is improved by 4.8% and 8.8%, confirming the model’s usefulness for crowd flow prediction tasks.

This paper proposes a multimodal human pose re-construction method based on 3D ultra-wideband (UWB) radar images and point clouds, aiming to improve the accuracy of human pose estimation through the fusion of radar images and point cloud data. First, a UWB 3D imaging radar system is designed, which synchronously collects radar echo signals and optical images, con-structing a multimodal dataset covering various common actions and different human characteristics. Radar data processing in-cludes azimuth-range 2D imaging, target locking, local 3D imag-ing, discrete sampling, and maximum projection to generate point cloud data and projection images. Optical image processing uses mature methods to reconstruct 3D poses as pose labels for point clouds and projection images. To achieve multimodal data fusion, the UWB FusionPose network is designed, comprising an image feature extraction network, a point cloud feature extraction net-work, and a pose reconstruction network. The image feature ex-traction network is based on the ResNet-18 framework, while the point cloud feature extraction network adopts a pyramid struc-ture. After feature fusion, a multi-layer perceptron (MLP) is used to predict human pose information. Additionally, this paper ex-plores the impact of fusion parameters on network performance and verifies the effectiveness of the multimodal network through ablation experiments. Experimental results show that this method effectively utilizes radar point cloud data and projection image data to accurately reconstruct the 3D pose of human targets. This research not only provides a new human pose reconstruction tech-nique but also offers valuable references for the future develop-ment of radar imaging technology and multimodal data fusion methods.