Improving the performance of the camshift algorithm using dynamic parallelism on GPU

The CamShift algorithm is widely used for tracking dynamically sized and positioned objects that appear in a sequence of video pictures captured by a camera. In spite of a great number of literatures regarding CamShift on the platform of CPU, its research on the massively parallel Graphics Processing Unit(GPU) platform is quite limited, where a GPU device is an emerging technology for high-performance computing. In this work, we improve the existing work by utilizing a new strategy – Dynamic Parallelism (DP), which helps to minimize the communication cost between a GPU device and the CPU. As far as we know, our project is the first proposal to utilize DP on a GPU device to further improve the CamShift algorithm. In experiments, we verify that our design is up to three times faster than the existing work due to applying DP, while we achieve the same tracking accuracy. These improvements allow the CamShift algorithm to be used in a more performance-demanding environment, for example, in real-time video processing with high-speed cameras or in processing videos with high resolution.