An improved method for vehicle tracking in UAV-captured video is introduced in this paper, this method based on GPU (Graphics Processing Units) acceleration which have efficient parallel computing powers. Today, vehicle tracking in UAV videos has many meaningful applications in computer vision, such as traffic management and illegal vehicle tracking. However, the increased computational complexity makes it difficult to be used in real-time tracking system. To overcome the limitation of tracking speed and meet the highly intensive calculation required, we perform image processing with GPUs and adopt an optimized structured output support vector machine (SVM) to online learning. The proposed approach is evaluated in UAV videos taken from DJI Matrice 100 drone and the corresponding conclusions are received.