Konstantinos Fatseas

University of Twente

EDL P 16-25 P7: Efficient Deep Learning Platforms (eDLP)

Research assignment
Deep learning for automotive radar sensors

Frequency Modulated Continuous Wave (FMCW) radar sensors have become essential for autonomous driving applications and driver assistance. The reason is that they can accurately measure the distance to targets and their radial velocity under adverse weather and lighting conditions. Furthermore, they are small, affordable, and ready for mass production.
Deep learning (DL) can assist in radar data processing, but a key obstacle for applying neural networks in the radar domain is the requirement for many annotated data samples to train them. For example, adding labels such as bounding boxes or segmentation masks to range-Doppler (RD) maps is a tedious and costly manual task. Furthermore, the RD maps can be hard to interpret because of clutter and ghost targets.
Due to the lack of datasets with raw radar data, we have developed a cost-efficient setup for data collection. The setup consists of a camera, an FMCW radar from NXP, and software to synchronize the two sensors and control the data collection.
Our work aims at developing DL-based methods to interpret the radar data we collect and extract the maximum possible amount of information about the objects in front of the radar. To do so, we focus on three areas: automatic radar data annotation, object detection in radar data, and camera-radar sensor fusion. These research fields are strongly interconnected to each other. For example, we can automatically generate the annotations to train a deep neural network for radar object detection using offline sensor-fusion methods.
Radar-camera setup for data collection.


Weekly Supervised Semantic Segmentation for Range-Doppler Maps. K. Fatseas and M.J.G. Bekooij. Conference proceedings at EURAD 2021. No open access.

Personal information