| Author: | Khan, Danista |
| Title: | Channel state information-based crowd counting and human activity recognition |
| Advisors: | Ho, Wang-hei Ivan (EEE) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Subject: | Human activity recognition Crowds Wireless sensor networks Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Electrical and Electronic Engineering |
| Pages: | xxiv, 132 pages : color illustrations |
| Language: | English |
| Abstract: | The widespread adoption of wireless technologies has introduced a new era of non-intrusive wireless sensing for human activity recognition and crowd counting. Wireless technologies like Wi-Fi allow us to seamlessly monitor and sense different aspects of our surroundings without invasive methods. This breakthrough has paved the way for applications of Internet-of-Things (IoT) making it effortless to collect data and track human activities. Techniques for device-free human activity recognition (HAR) based on radio-frequency (RF) are divided into received signal strength indicator (RSSI) and channel state information (CSI)-based approaches. RSSI-based techniques rely on path loss per packet to sense human activities. In contrast, CSI-based methods capture fine-grained data that contains the combined effect on the wireless signals by the dynamics in the surroundings. Although it is easier to collect RSSI measurements, CSI-based HAR is preferred as it contains the effect of multipath propagation by human motion which produces a unique pattern in the time series and helps to classify human activities accurately. Device deployment is crucial in wireless sensing, ensuring optimal wireless coverage for the intended sensing application. The sensing boundary of a wireless link relies heavily on the line-of-sight (LOS) distance between the transmitter and receiver. In this thesis, for the first time in literature, we conduct a series of experiments to demonstrate the impact of LOS distance on the classification of the number of people in an indoor environment. As the distance increases, the classification accuracy initially improves due to the increase in sensing coverage area, peaking at 3.5 meters. However, beyond this point, the accuracy declines as the distance continues to increase. Our experimental findings demonstrate that leveraging LOS path distance can effectively control the wireless sensing boundaries for multiple targets, which helps mitigate interference issues in people counting for the indoor environment. Our experimental results provide insights into the scalable deployment of wireless sensing applications, with multiple wireless links, ensuring high precision and accuracy in diverse environments. In this thesis, we investigate CSI-based HAR and crowd counting to provide a comprehensive study on how CSI data is collected to generate unique patterns to accurately classify different human activities and the number of people in an indoor environment by exploiting different deep learning models. Prior studies on activity recognition mostly differentiate human activities by classifying one complete series into an activity. However, these approaches require massive datasets to give accurate results in real-time scenarios, and the classification is in fact based on short-term activity samples instead of the complete activity series. In this thesis, we implement sample-level activity recognition by exploiting a special type of convolutional neural network (CNN), U-Net, which can utilize multi-layer architecture to learn features automatically and classify them on the pixel level. Hence, this neural network can classify human activities based on sample level by treating the sample as a pixel of the CSI frame. The data collection setup does not require manual feature extraction and can efficiently classify short-term activity samples with an average accuracy of 98.57%. State-of-the-art CSI-based supervised crowd counting systems are vulnerable to temporal and environmental dynamics in practical scenarios as their performance degrades with fluctuations in indoor environments due to multipath fading. Inspired by the breakthroughs of transfer learning and advancement in edge computing, in this thesis, we have leveraged the concept of transfer learning to minimize the problem of temporal and environmental dynamics. The trained model from the source environment is exploited for other indoor environments to perform device-free crowd counting (CrossCount) at the target rooms. Our results show that this technique can combat the dynamics of the environment and achieves 4. 7% better accuracy with a 40% reduction in training time compared to conventional CNN. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13184