Towards Multi-Object-Tracking with Radar on a Fast Moving Vehicle: On the Potential of Processing Radar in the Frequency Domain

A team of researchers including Tim Hansen, Arturo Gomez-Chavez, Ilya Shimchik, and Andreas Birk has published a paper advocating for a fundamental shift in how radar data is processed for autonomous vehicles. Their research, titled "Towards Multi-Object-Tracking with Radar on a Fast Moving Vehicle: On the Potential of Processing Radar in the Frequency Domain," argues that moving from traditional feature-based methods to frequency-domain analysis offers superior robustness. This approach is particularly crucial in high-dynamic environments where a vehicle's own motion (ego-motion) combines with multiple other moving objects, creating a complex and noisy scene. The frequency-domain processing provides inherent noise resistance and, as a key advantage, captures information about all moving structures simultaneously through underlying correlation methods used for tasks like registration.

To validate their claims, the researchers developed and tested an initial implementation called Fourier SOFT in 2D (FS2D). They demonstrated its capability for radar-only odometry—calculating a vehicle's movement using radar alone, without fusing data from other sensors like cameras or lidar. This proof-of-concept was tested using the Boreas dataset, a benchmark for all-weather autonomous driving research. The paper uses the demanding scenario of an autonomous racing overtaking maneuver as a motivating example, highlighting the method's potential to handle the extreme speeds and close-quarters tracking required. This work, available on arXiv under identifier 2604.14013, represents a significant exploration into making radar a more reliable and standalone perception modality for next-generation autonomy.