Jelena Senic’s research while affiliated with University of Colorado Boulder and other places

We describe a quasi-deterministic channel propagation model for human gesture recognition reduced from real-time measurements with our context-aware channel sounder, considering four human subjects and 20 distinct body motions, for a total of 120 000 channel acquisitions. The sounder features a radio-frequency (RF) system with 28 GHz phased-array antennas to extract discrete multipaths backscattered from the body in path gain, delay, azimuth angle-of-arrival, and elevation angle-of-arrival domains, and features camera / Lidar systems to extract discrete keypoints that correspond to salient parts of the body in the same domains as the multipaths. Thanks to the precision of the RF system, with average error of only 0.1 ns in delay and 0.2∘ in angle, we can reliably associate the multipaths to the keypoints. This enables modeling the backscatter properties of individual body parts, such as Radar cross-section and correlation time. Once the model is reduced from the measurements, the channel is realized through raytracing a stickman of keypoints – the deterministic component of the model to represent generalizable motion – superimposed with a Sum-of-Sinusoids process – the stochastic component of the model to render enhanced accuracy. Finally, the channel realizations are compared to the measurements, substantiating the model’s high fidelity.

Millimeter-wave (MmWave) channel characteristics are quite different from sub-6 GHz frequency bands. The major differences include higher path loss and sparser multipath components (MPCs), resulting in more significant time-varying characteristics in mmWave channels. It is difficult to characterize the time-varying characteristics of mmWave channels through statistical models, e.g. slope-intercept models for path loss and lognormal models for delay spread and angular spread. Therefore, highly accurate channel modeling and prediction are necessary for deployment of mmWave communication systems. In this paper, a mmWave channel parameter prediction method using deep learning and environment point cloud is proposed. The parameters predicted include path loss, root-mean-square (RMS) delay spread, angular spread and Rician K factor. First, we introduce a novel measurement campaign for indoor mmWave channel at 60 GHz, where a light detection and ranging (LiDAR) sensor and panoramic camera were co-located with a channel sounder and then time-synchronized point clouds and images were captured to describe environmental information. Furthermore, a fusion method between the point clouds and images is proposed based on geometric relationship between the LiDAR and camera, to compress the size of the data collected. Second, based on a classic point cloud classification model (PointNet), we propose a novel regression PointNet model applied to channel parameter prediction. To overcome generalization problem of model under limited measurements, an area-by-area training and testing method is proposed. Third, we evaluate the proposed prediction model and training method, by comparing prediction results with measured ground truth. To provide insights on what training inputs are best, we demonstrate the impacts of different combinations of input information on prediction accuracy. Last, the deployment and implementation method of the proposed model is recommended to the readers.

Accurate channel propagation modeling of foliage is critical to the design of wireless networks, given its pervasive nature in rural, suburban, and urban environments. Its blockage effects can be particularly devastating at millimeter-wave (mmWave) because the size of leaves and branches is comparable to the wavelength of the transmitted signal. While raytracing models are firmly based on electromagnetic principles, reliability can be attained only through calibration against measurements. In the few works that do so, foliage is represented as simple canonical shapes (cylinders, discs, etc.) and calibration is performed against measurements with foliage integrated as part of entire outdoor environments. The controlled approach that we adopt in this paper, rather, is based on measurement of single specimens of foliage, for precision characterization. To sustain this precision at mmWave, the foliage is represented digitally as a mesh of faceted leaves and branches. Raytracing predictions from the Ansys HFSS SBR+ model applied to digital twins of seven trees are calibrated against measurements – collected in summer and in winter for comprehensive analysis – with the Terragraph double-directional 60 GHz channel sounder. The tree-specific predictions, which can then be integrated as part of an entire outdoor environment, are shown to match the measurements very well.

There is strong impetus by the Telecom Infra Project to exploit the 60 GHz unlicensed band for public Wi-Fi in urban environments, by installing access points on light poles. Although many 60 GHz urban channel measurements have been recorded to date, they have resulted only in path loss models or root-mean-square (RMS) delay spreads. What is needed at millimeter-wave is a spatially consistent channel model for beamtracking that embodies the characteristics of these short wavelengths—sparsity and rough surface scattering—such as the Quasi-Deterministic model. In this letter, we fit the model to channel measurements we recorded in an urban environment. The measurements were recorded at 4, 6, and 9 m antenna heights to investigate the tradeoffs between light pole heights. The large-scale channel metrics between the model and the measurements were shown to match very well.

Human blockage at millimeter-wave frequencies is most commonly modeled through Knife-Edge Diffraction (KED) from the edges of the body shaped as a vertical strip. Although extensively validated in controlled laboratory experiments, the model does not scale to realistic 3D scenarios containing multiple, randomly oriented human blockers, on which multipath signals can be incident from any direction. To address this, in this article we investigate numerical approaches based on ray-tracing methods. Predictions from two electromagnetic computational methods in addition to the KED, namely the Uniform Theory of Diffraction (UTD) and Physical Optics (PO), are compared to an extensive suite of precision measurements at 60 GHz. Besides the vertical strip, cylinder and hexagon body shapes are considered with the UTD method, and a 3D phantom shape is considered with the PO method. We found that the PO method is the most accurate, but also the most computationally intensive due to the large number of faces (approximately 8000) in the phantom and due to the inherent complexity of the method itself. While the UTD method with the hexagon shape (approximately 42 faces) is slightly less accurate than the PO method, it provides the best compromise when efficiency is paramount.