Ultra-wideband Radar Child Presence Detection Using Deep Learning
- In work
- Master Thesis
- Announcement date
- 14 Feb 2022
- Research Areas
The extremely high or low temperatures that can develop in parked cars are a threat to children left in those cars. To avoid this threat, and thereby potentially save lives, we need to detect if children are present in cars when they are being locked.
Several systems to achieve this task are proposed, e.g. by using cameras or weight sensors. However, those systems have drawbacks and limitations, such as depending on illumination conditions, or not being able to distinguish between lifeless objects such as a suitcase or a bag of groceries and living beings. In addition, those systems often require car manufacturers to include dedicated sensors for this task which increases manufacturing complexity and cost or raises privacy concerns. Many of these problems can be avoided by using ultra-Wideband radar sensors. These nodes are already placed in the car for the keyless-entry feature and thus do not incur additional cost to the manufacturer. This makes it even possible to retrofit such a child presence detection algorithm to existing cars that already possess the required UWB nodes with a software update.
Nevertheless, detecting the presence of a person by means of the captured radar signal is not a trivial task, due to the strong multipath components and received clutter inside the car. In order to separate the clutter form the target signal, we currently focus on detecting the breathing motion of the chest of the target. Furthermore, the decision should be available within 10s after the car being locked. This includes measurement time as well as the runtime of the algorithm. Therefore, the runtime of the algorithm should be as short as possible to allow for a longer measurement time.
- Research literature on machine learning for UWB occupancy detection and how to adapt algorithms to generalize well when trained on simulated data.
- Design and implement a machine learning architecture (preferably in python) and train the algorithm using simulated data.
- Use transfer learning on a very small real-world dataset to ensure that the trained network generalizes well to the real world.
- Evaluate the performance of the developed algorithm and compare the performance against the existing classical detection algorithm.
What we expect from you:
- You are master student in information and computer engineering, electrical engineering, audio engineering or similar studies.
- You are interested in and know the basics of machine learning / deep learning as well as signal processing.
- You have experience in programming in python, matlab or some other science-oriented programming language.
- Optional: You already have some experience in designing and implementing machine learning algorithms using PyTorch, Tensorflow or a similar framework.
If you are interested and want to know more about it, send me an email to firstname.lastname@example.org or visit me in my office at Inffeldgasse 16c room IDEG006.
This thesis is in cooperation with NXP Semiconductors. A part-time employment in the project will be offered.