WiFi channel state information (CSI) and machine learning-based environmental sensing have gained significant attention. However, it is empirically known that the accuracy of trained models degrades over time. In this study, we experimentally demonstrate that this temporal degradation in CSI sensing model accuracy is caused by concept drift, a phenomenon where the distribution of CSI shifts over time. Specifically, we provide evidence of long-term concept drift based on CSI measurements taken in a static environment over the course of seven days. Furthermore, we show that by applying noise-based data augmentation during model training, the impact of long-term concept drift can be mitigated. This approach alleviates the need for continuous model fine-tuning, which is resource-intensive due to its reliance on costly labeled data.

We introduce CSI-Inpainter, a novel approach for obstacle removal using Wi-Fi CSI. This method harnesses CSI data to reconstruct obscured visual elements, regardless of lighting conditions. Extensive empirical evaluation in both office and industrial settings demonstrates the effectiveness of CSI-Inpainter’s exceptional ability to identify and reconstruct occluded segments, outperforming traditional baselines and our received signal strength indicator (RSSI)-based work, RF-Inpainter in terms of visual quality. Our findings emphasize the superiority of CSI data over RSSI for providing richer visual information and underscore the critical role of optimal sensor placement and data fusion from multiple CSI sensors in enhancing the performance. CSI-Inpainter represents a significant advancement in obstacle removal for various applications like surveillance, offering new insights into the integration of wireless sensing and visual scene recovery, expanding the potential applications of Computer Vision in real-world environments.

Split computing (SC) introduces an innovative framework designed for inference operations of deep neural networks (DNNs), enhancing traditional machine learning inference approaches that typically deploy models solely on local devices or edge/cloud servers. This research not only implements the SC framework for distributed machine learning inference but also showcases its application in computer vision tasks. We demonstrate the capability of SC to execute privacy-preserving distributed inferences across both image classification and object detection tasks. Our comprehensive implementation will be made publicly available at https://github.com/nishio-laboratory/.

This paper demonstrates the feasibility of millimeter-wave (mmWave) link quality prediction based on out-of-band channel state information (CSI) fingerprinting. To overcome the pedestrian blockage problem of mmWave communications, a large number of computer vision-aided mmWave link quality prediction methods have been investigated. However, the use of cameras and LiDAR to acquire computer vision information entails privacy risks. In this paper, we employ 5 GHz band CSI fingerprinting - an aggregation of CSI measured at multiple locations, for mmWave link quality prediction. CSI reflects the propagation environment of the wireless communication channel and thus includes information on pedestrians that block mmWave communications. CSI fingerprinting, aggregated from various measurement locations, enables future mmWave link quality prediction owing to its sufficient spatial information. We conducted a real-world wireless communication experiment with commercial devices compliant with IEEE 802.11ad for the mmWave, and nine IEEE 802.11ac CSI measurement devices for 5 GHz, to experimentally evaluate our method. The experimental result revealed that our proposed method can deterministically and numerically predict the mmWave link quality deterioration caused by pedestrian blockage 500 ms in advance.

This paper proposes a feature extraction method from three-dimensional (3D) point cloud employing spatial frequency analysis for point cloud-based link quality prediction. The proposed method aims to address the human blockage problem in millimeter-wave (mmWave) communications, where proactive communication control through machine learning-based future link quality prediction has been shown to be effective in preventing link quality degradation. However, the use of 3D point clouds presents challenges such as increased transmission cost and computational complexity due to the large data size. To address these challenges, we propose a preprocessing method that can efficiently extract features from the point cloud and reduce data size without compromising the accuracy of mmWave link quality prediction. Our approach uses spatial frequency-based filtering to isolate signals related to moving obstacles, such as pedestrians, in the frequency domain. It also removes large, static background objects like walls and furniture. The proposed method can extract relevant objects in the point cloud based on their spatial frequency without requiring object detection or segmentation. The experimental results demonstrate that our proposed method significantly reduces feature data size by approximately 99% compared to conventional methods, while still maintaining high link quality prediction accuracy.

This paper demonstrates the feasibility of predicting the future received power of millimeter-wave (mmWave) communication using point cloud data. To mitigate the human blockage problem in mmWave communication, previous works have studied a camera-vision assisted mmWave link quality prediction, which predicts the time-series of the received power from the next moment to as many as several hundred milliseconds ahead by leveraging camera imagery and machine learning. However, camera imagery generally includes privacy-sensitive information, which induces privacy concerns in the camera-assisted mmWave networks. In this paper, we demonstrate that point cloud, which can be obtained by light detection and ranging (LiDAR) sensors and poses fewer privacy concerns than camera imagery, can be an alternative to the cameras in mmWave link quality prediction. Specifically, we propose a mmWave received power prediction method using point cloud data, and our experimental evaluation demonstrates that the proposed method predicts mmWave received power 500ms ahead with a root-mean-squared error of 3.3dB, which is comparable to the existing camera-based method. Moreover, we verify that even minor environmental changes can degrade the accuracy of the trained prediction model and this degradation can be mitigated by model fine-tuning with a small dataset.