Comparing Data Efficiencies in Vision Aided Wireless Communications
Sean Teng
Gpbs internet speeds with the help of cameras, deep learning and millimetre wave!
Millimetre-Wave mobile communication can generate Gpbs speeds for you! However, the main challenges of implementing mobile communication at millimetre Wave frequencies lie in the physics of millimetre Wave propagation. Compared to sub-6GHz propagation, the frequencies that our WiFi and cellular connections are currently on, millimetre Wave struggles with blockages in the propagation path because of high path and penetration losses, and poor diffraction. Therefore, signal blockages from common objects, such as human bodies, can cause signal degradation and render millimetre wave unviable.
However, we can use deep learning to predict better directions of signal propagation, estimate millimetre wave channels and detect blockages. These predictions using deep learning can make millimetre wave mobile communications a reality by overcoming the obstacles caused by millimetre wave propagation. Moreover, there is a field of research focused on using cameras and other visual sensors, for these deep learning predictions. This field is called Vision Aided Wireless Communications.
My project aims to discover the advantages and disadvantages of incorporating data from visual sensors, by comparing the computational efficiencies of different data types, such as RGB, LiDAR and wireless data!
Page Views:
In order to compare data modalities as objectively as possible, the project consists of designing a neural network architecture that is relatively agnostic to any data modality. In other words, apart from the input stage of the network, where the data is processed, the rest of the architecture should remain consistent. In addition, the model architecture should be as close to an architecture seen in production as possible. Therefore, the proposed model should be able to perform multi-task learning. Thereby performing three tasks of beam, blockage and channel prediction, the three key problems in millimetre wave mobile communications. Furthermore, the inference performance of the data modailities will be compared on a desktop GPU and an Nvidia Jetson Nano, to benchmark the performance drop experienced when deploying models in production. As far as I know, the conclusions made from this FYP will be new findings in the field of deep learning in millimetre wave mobile communications.
If you would like to learn more, I have written a survey paper on this topic, which is currently under peer review with IEEE COMST. Please stay tuned at my ORCID https://orcid.org/0000-0002-1667-1074, where I will link my paper once it is published. Thank you!
