Development of a WiFI based human activity detection and recognition system

• 主要作者: Ji, Junwei
• 其他作者: Xie Lihua
• 格式: Thesis-Master by Coursework
• 語言: English
• 出版: Nanyang Technological University 2021
• 主題: Engineering::Electrical and electronic engineering::Electronic systems::Signal processing
• 在線閱讀: https://hdl.handle.net/10356/154346

In today's era, with the rapid development of information technology, people are increasingly inseparable from WIFI. WIFI seems to have become an indispensable part of our lives. Research on WIFI has also attracted more and more attention from the industry. Among them, WIFI-based human behavior recognition has gradually become a hot spot in the research of the Internet of Things. Human behavior recognition (HBR) is very useful in our daily lives, and its application scenarios are also very wide, such as early warning reminders, intelligent security, patient monitoring systems, etc. Therefore, it is of great significance to study different methods of human behavior recognition. The most well-known method is machine vision, which collects video information through a camera system, and then uses image processing related algorithms to realize behavior recognition. Another method is to allow the detected person to wear sensors. This method mainly detects the subtle activity information of the human body, such as changes in joints or fingers. Although this method can guarantee accuracy, there are problems such as personal privacy, system complexity, and expensive equipment. Therefore, there are a lot of limitations in practical use. Thus, the wireless HBR system is proposed, such as ultra-wideband (UWB) based human behavior recognition. Since WIFI is more common in daily life, human behavior recognition based on WIFI has been paid more attention, and more and more scholars have begun to study it. So far, there are three main methods for studying this subject: model-based, pattern-based and neural network-based. This dissertation chooses to use neural network to study the subject and improves the Autoencoder Long-term Recurrent Convolutional Networks (AE-LRCN) algorithm. The discrete Wavelet transform (DWT) algorithm is used to reduce the dimensionality of the data to obtain two coefficients which will be put into the autoencoder. Then use the Canonical Correlation Analysis (CCA) algorithm to fuse the two sets of feature vectors obtained from the autoencoder. Finally, LRCN algorithm is used to realize human behavior recognition. The proposed algorithm further shortens the training time and improves the convergence speed of the neural network while maintaining the high accuracy of the original algorithm. According to experiments, it can be found that no matter what kind of clothes a person wears (long-sleeved and short-sleeved), it does not affect the recognition of human behavior, and the accuracy rate can reach nearly 99%.