My master thesis gives a detailed insight into bundle adjustment as a powerful tool for globally optimizing maps created by feature based simultaneous localization and mapping (SLAM). Background knowledge, especially concerning the mathematical theory, is illustrated clearly. Different problems affecting robustness and efficiency are examined and solutions are given. Therefore a major part focuses on the parallel implementation in OpenCL, a language for parallelizing algorithms for different processor architectures like CPUs, GPUs and FPGAs. The algorithm is tested against different synthetic and real datasets and the results are shown graphically using Point Clouds and the CVT library. In addition, different implementations are compared with each other concerning robustness and performance. It is illustrated that a hybrid OpenCL bundle adjustment implementation, which is using a GPU and a CPU simultaneously, outperforms all other approaches.