Query-Centric Radar Perception via Structured Axes of Reasoning
The doctoral thesis of Dr.-Ing. Loveneet Saini investigates how the specific characteristics of radar data can be leveraged to enable reliable vehicle perception. To this end, he develops a query-centric approach that captures spatial, temporal, and graph-based structures in radar data, which is often sparse and noisy, and integrates them within a unified architecture. The proposed methods improve tasks such as object detection, velocity estimation, and segmentation on both public and industrial datasets.
We asked Saini about his dissertation:
In what context was your dissertation written? Which projects or other factors particularly influenced your dissertation?
My dissertation was developed as an industrial PhD in cooperation with Aptiv, where I worked on radar-based perception for automated driving. Working daily with real sensor data and production requirements shaped the work from the start. The driving question was how transformer architectures, which transformed camera-based perception, could be adapted to radar data, which is sparse and noisy but carries unique velocity information. From this tension between research and application, the central idea emerged: a query-centric view of radar perception.
What contribution does your work make to the field of research?
Radar perception has long stood in the shadow of camera- and lidar-based research, despite radar's robustness and cost efficiency making it essential in the automotive industry. My work shows that query-based transformer architectures, structured along spatial, temporal, and relational axes of reasoning, can significantly improve radar object detection in bird's-eye view. The approach was evaluated on the nuScenes benchmark, published at WACV
and CVPR workshops.
What's next for you and the topic?
I recently started as a Senior Engineer for Autonomy & AI at ARX Robotics in Munich, working on perception and autonomy for unmanned ground vehicles. Many questions from my dissertation carry over directly, such as robust perception under difficult conditions and efficient architectures on embedded hardware. AI perception remains an open and exciting field, and I look forward to continuing to bridge research and real-world autonomous systems.
