Abstract
In the past few years, researchers have worked on using Wi-Fi for identifying actions, but this overlooked using Wi-Fi on static (non-moving) or relatively non-moving objects due to its inherent problems. These problems involve many challenges such as very close range, high clutter, running real time, rough blob-like images, and many other issues encountered. Another area where both have had little research is in classifying the images created with Wi-Fi. Machine learning is a growing topic due to fast GPUs, but no research has been conducted in applying artificial intelligence (AI) to classifying non-moving or relatively non-moving objects imaged with Wi-Fi signals. This research used Wi-Fi signal processing and neural networks to identify static objects. The images created needed to be classified (identified). Classifying the resultant images were a challenge and required some form of AI. AI can identify objects by seeing persistent characteristics in rough images. Here is where Wi-Fi imaging and AI worked hand in hand. There were two major portions to this research. The first was the signal processing portion, where images were created, and the second was the image classification portion, which used AI neural net-works to identify the rough images. This work brought the researcher’s theoretical ideas to life in real-world experiments, which provided the proof of concept. In the signal processing portion, images were created by using Wi-Fi signals transmitted and received on Ettus Universal Software Radio Peripheral (USRP) hardware and directional antennas. Different directional antennas were designed and manufactured to form 1x2, 2x2, and 3x2 antenna arrays for beamforming and scanning objects. Upon creating Wi-Fi images, region-based convolutional neural networks(RCNNs) were employed and trained to result in identifying these objects. The results showed that such a Wi-Fi imaging and classification system is possible for static objects. This research solved problems of identifying static objects not well seen by visible light cameras, night vision, thermal vision, or millimeter wavelength processing where cost or environment factors interfered. Problem areas that were shown to benefit from this research were fire and rescue robotics, automation of security in bus and subway systems, and many other applications.