Parallelization analysis of ECO tracking algorithm on GPUs

Object tracking is a very popular area in image processing. Its popularity comes from the variety of its application areas. It is used for security and surveillance, autonomous vehicles, human-machine interaction, traffic control and so on. Due to its application areas, an object tracking algorithm is usually expected to be fast. On the other hand, an object tracking algorithm should be accurate and robust and this usually increase the amount of calculations to be done. The nature of the many image processing applications are suitable for parallel programming. Since, GPUs consist of large number cores, they are widely used in image processing and object tracking applications. In this thesis, we analyze an object tracking algorithm for its suitability of parallelism. We detected the time-consuming parts of the algorithm by using profiling tool. Each part of the algorithm is handled separately and implemented on GPU. Additionally, we have worked on the chances of optimization by using GPU capabilities. We compared our methods with the original parts of CPU based approach by testing them on five datasets.