Parallel Optimization of Correlation Filter Tracking Based on CUDA

As one of the most important research directions in computer vision, real-time video object tracking technology based on online learning has been widely applied in various fields. Factors such as deformation, occlusion, and similar background interference can seriously affect the performance of the tracker, and slow computation can also significantly reduce the real-time performance of the tracker. In order to solve the problem of slow running speed and difficulty in meeting real-time performance of tracking algorithms on low-cost computing platforms, this paper conducts a high-performance implementation of correlation filter tracking based on CUDA parallel programming technology and a heterogeneous architecture of CPU+GPU. It focuses on parallel acceleration of feature extraction, fast Fourier transform, and kernel correlation calculation. On this basis, the optimized algorithm is deployed on the low-cost embedded platform. The experiment proves that the average acceleration ratio of the optimized tracking algorithm reached 67.5%, fully meeting the tracking requirements in actual engineering project scenarios. Finally, based on the research results of this article, a high-speed real-time target tracking system is built, and the actual application testing of the overall algorithm is completed.