Tracking of Highly Dynamic Objects

• Applicant: Professor Dr. Andreas Birk
• Subject Area: Traffic and Transport Systems, Intelligent and Automated Traffic
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 561031914

The tracking of moving objects from a moving platform is a well investigated but nonetheless also still very challenging problem, especially in scenarios with high requirements regarding the robustness and the processing speed like in autonomous driving. Two major aspects need to be considered: [A] the motion of the sensor(s), i.e., the estimation of the ego-motion of the platform itself and [B] the detection and tracking of moving objects in the sensor-data, which is particularly challenging when there are an unknown number of objects, high dynamics of the objects, and also sensor noise. State of the art approaches tend to try to solve [A] first to tackle [B] second in ego-motion-compensated data. In the proposed project, the idea is to use spectral registration methods to investigate among others the options and also limitations of addressing both aspects in parallel. Spectral registration methods in 2D and 3D operate in the frequency domain. As it is well-known since the early days of computer vision, translation parameters can be determined by phase correlation using the Fast Fourier Transform (FFT). It is much more challenging to also determine rotation. For this, approaches exist using in addition, e.g., the Mellin-Transform or the SO3-Fourier-Transform. Spectral registration methods generate in theory a Dirac pulse, respectively a broader peak in practice when using real-world data with noise, occlusions, dynamics, etc., for each degree of freedom of the relative motion-parameters of the sensor-poses, i.e., for the ego-motion. In addition, peaks are generated for each moving structure in the sensor data, i.e., for the to be tracked objects. A major advantage of using spectral methods is hence that the number of tracked objects does not need to be known a priori. Also, there is no need for using any involved object detection method, which would require additional computation. Further relevant advantages of spectral methods are that they have a fixed, fast computation time, that they are well-suited for parallelization, and that they tend to be robust against various forms of noise. Autonomous driving at high velocities, i.e., autonomous racing with full-scale racecars, is used as investigation and demonstration scenario for this research, especially in the context of over-taking maneuvers where other racecars at high velocities need to be detected and tracked. The available full-scale racecar used to this end features multiple lidar and radar sensors, for which a core objective is to process their data on the vehicle itself in real-time under competition conditions, i.e., while racing against other cars on a Formula-1 race-track.

• DFG Programme: Research Grants