High-Speed Tracking of Discriminative Correlation Filters Based on GPU

Most existing high-speed trackers are based on generative models, which cannot distinguish between target and background. In contrast, the discriminative correlation filter (DCF) based on discriminative model treats tracking as a classification of background and target. To address the problem that DCF cannot track at high speed. In this paper, we propose a high-speed GPU-based parallel discriminative correlation filter (PDCF), decompose the DCF into parallel tasks and map them to GPU with different coarse-grained and multi-level memory. The experiments result show that PDCF is 58 times faster than DCF running on a single CPU with the same accuracy, and the speed is 2213 frames per second (FPS) when the patch size is 200*200pixel, which fully satisfies the task of high-speed tracking.