Hochauflösende Prozessierung flugzeuggestützter SAR Daten auf GPU

Kurzfassung

The airborne based radar system F-SAR built and operated by the Microwave and Radar Systems Institute is a major contribution to the project VABENE. With its help it is possible to take pictures of the earth independent from weather and daylight. In contrast to optical pictures it is necessary to transform the incoming signals using computationally expensive algorithms in the first place, so that humans can interprete them. Within the work of the VABENE project there is the need to acquire high resolution data as fast as possible. Neither the possibility to compute radar images on board nor offline on the ground meets the requirements. The onboard processor only has low resolution and full resolution can’t be provided in real-time. That’s the reason why CPU-based hardware architectures will not allow solving this problem. New technologies are needed and that’s why general purpose graphics processing units (GPGPU) are playing an important role. Normal GPUs are specialized to work with computergraphics. With the help of the programming architecture CUDA they become a computing machine capable of solving computationally demanding tasks in short time. GPUs are normally used for computer graphics because they are designed for massiv parallelized computations. The development of the last years shows that, because of the programming architecture CUDA, it is now possible to port radar data processing on GPU to speed up processing time. The topic of this bachelor thesis is to implement the radar data proccessing steps for GPU usage and to find a way to integrate the software into the existing software systems of the F-SAR processor. Tests show that integrating CUDA code in the existing software architectures is best done by implementing dynamic libraries – shared objects. With these, CUDA code is portable and can be used by every C or IDL program. It is also important to be flexible as algorithmic steps in the radar data processing chain may change. Benchmark tests performed for every module have shown that the computing time of individual modules compared to single-core CPU processing is 10 to 70 times faster. That shows clearly the potential of the GPU architecture. But there are more possibilities left to further optimize the proccessing for full real-time processing. Because the algorithms for radar data processing can be massively parallelized, the capacity of the GPU computation force can be used efficently. Therefore porting the whole radar processing to the GPU enables the radar system to calculate a full resolution radar image in real time despite of the complex algorithms used. However, because of present F-SAR hardware restrictions this performance is presently limited to offline environments.