Compressed generative adversarial networks for real-time feature extraction for fingerprinting localization with smartphones.
Today smartphones are not only a mean of communication but also a source of rich information for a variety of sensing applications from entertainments, social media, to healthcare. However, the sensor information is meaningless without location information. In addition, people carrying smartphones most of the time. Therefore, tracking smartphones in real-time is essential for many applications. Among the localization techniques, radio frequency fingerprinting is the most popular as every smartphone is equipped with a WiFi/Bluetooth interface. However, the received signal strength by itself is too noisy to obtain a good localization estimation. Recent development in Generative Adversarial Networks provides a promising way to extract more useful feature from raw received signal strength to improve the localization performance.
The goal is to develop a compressed GAN and to implement it on smartphones, which can read and extract useful hidden feature from a stream of raw received signal strengths in real-time to enhance fingerprinting localization.
The project will be divided into 3 main tasks:
- Developing a tunable, light-weight model that can classify ECGs, given a pre-trained model
- Implement the tunable model on a Raspberry Pi Zero
- Examining both the accuracy and efficiency of the implemented model.
20% Theory, 60% Implementation, 20%Writing
Le Viet Duc, firstname.lastname@example.org, room ZI 5013