ORCID Identifier(s)


Graduation Semester and Year




Document Type


Degree Name

Doctor of Philosophy in Computer Science


Computer Science and Engineering

First Advisor

Yonghe Liu


With the rapid development of 802.11 standard and Internet of Things (IoT) applications, Wi-Fi (IEEE 802.11) has emerged as the most widely used wireless communication technology. Wi-Fi based sensing has found widespread use cases involving activity recognition, indoor localization, design of smart spaces and in healthcare applications. This dissertation presents the study of human activities’ sensing and recognition using channel state information (CSI) of Wi-Fi. We highlight the limitations of existing methods and consequently design the frameworks for collecting stable CSI and monitoring different indoor and outdoor environments for human activities. Specifically, this dissertation provide means to define and extract quality features in different noisy environments for achieving accurate human activity recognition (HAR). In first part of the dissertation, we present WiChase framework which extracts anomalous CSI data and processes it to sense and classify surrounding human activities. Contrary to some other works that use fixed CSI window, we automate the data extraction and activity labeling based on the variance of multiple input and multiple-output (MIMO) subcarriers. Moreover, we recognize human activities using CSI amplitude together with calibrated phase and propose the integration of these two factors for obtaining high accuracy. Importantly, we identified that exploiting MIMO subcarrier classification and consequently performing packet level majority voting of subcarriers significantly improves the accuracy. We define multiple sophisticated features for both amplitude and phase for our system and use multiple machine learning algorithms for the classification of human activities included in our evaluation. We then identified the limitations of CSI based sensing systems; requiring stable and noise-free environments. In our second work, we consider this and present a CSI based driver activity recognition framework for an in-vehicular environment which is inherently more noisy and unstable in contrast to indoor spaces like offices, experimental labs etc. For this research, we propose and design the first Wi-Fi based driver state recognition system: SafeDrive-Fi. Our proposed work extracts CSI of Wi-Fi to accurately predict driver states through gestures and body movements. Different from vision-based techniques, SafeDriveFi provides a simple, cost-effective and ubiquitous solution to prevent accidents and loss of lives due to reckless driving. We incorporate different signal processing techniques to differentiate between normal and dangerous driving in a challenging and noisy in-vehicle conditions. To improve driver states classification accuracy, we propose multi-level filtering together with subcarrier majority voting. Using today’s commercially available products,SafeDrive-Fi is compatible with 802.11n/ac and can assist drivers and law enforcement in discovering dangerous driving states. To the best of our knowledge, this is the first system that aggregates information from all the channel subcarriers and use multi-domain CSI features to classify dangerous driving conditions effectively. This part of dissertation also considers different parametric variations and other physical layer (PHY) information like Received Signal Strength Indicator (RSSI) for comparison purposes. In our evaluation of the proposed system, we varied diverse parameters like number of training samples, number of MIMO subcarriers, number of Tx-Rx streams etc., to analyze the system’s accuracy from multiple perspectives. The last part of this dissertation is focused towards extracting automated features from CSI containing human activities based perturbations. To achieve this, we propose a methodology to improve features extraction and learning in an effort to accurately classify multiple human activities. Existing works on human activity recognition predominantly consider single-person scenarios, which deviates significantly from real world where multiple people exist simultaneously. In this research, we leverage transfer learning, a deep learning technique, to present a framework (TL-HAR) that accurately detects multiple human activities; exploiting CSI of WiFi extracted from 802.11n. Specifically, for the first time we employ packet-level classification and image transformation together with transfer learning to classify complex scenario of multiple human activities. TL-HAR transforms CSI to images to capture correlation among subcarriers and use a deep Convolutional Neural Network (d-CNN) to extract representative features for the classification. We further reduce training complexity through transfer learning, that infers knowledge from a pretrained model. The detailed experimental results confirm the significance of optimal features extraction achieved through domain transformation and deep neural networks.We also present a 3D CNN based human activity recognition framework. CSI is acomplex quantity; however most of the researches have either utilized only amplitude of CSI or otherwise considered phase independently. Doing this not only retards efficient feature learning for a CNN but also loses significant correlation between amplitude and phase. In this work, we show that this correlation between amplitude and phase of CSI is important and carries significant activities information. Therefore, we present a methodology of incorporating both amplitude and phase of CSI simultaneously for extracting quality features to obtain high classification accuracy. Our results show that 3D CNN performs better andachieves 96.45% accuracy in contrast to 2D CNN which only obtains maximum accuracyof 94.20% even after using twice training samples in comparison to 3D CNN.


Human activity recognition, Sensor-less sensing, Channel State Information (CSI), Wi-Fi sensing, Machine learning, Deep learning


Computer Sciences | Physical Sciences and Mathematics


Degree granted by The University of Texas at Arlington