The Research Target Tracking Algorithm Based On GPU

With the rapid development of computer vision,video-based object tracking has become an important research direction, and have made great progress in recent years. New target tracking algorithms have been put forward, and the accuracy of tracking has been improved. However, a large number of experiments show that algorithms with high accuracy also have high computational complexity, so low processing speed becomes the one of the bottleneck. GPU provided by NVIDIA corporation has been successfully used in various fields owing to its powerful computing ability, and it provides a new solution to the complex algorithms with high parallelism and multi-intensive computing. Based on the above background, this paper do in-depth studies on achieving three algorithms—DF tracking algorithm, Struck tracking algorithm and KCF tracking algorithm, by using GPU. As for each kind of tracking algorithm, first doing detailed analysis on model training and target identifying to identify parallel computing sections. Then, use GPU to accelerate these sections. Finally, compare CPU implementation with GPU implementation. Experimental data shows that the implementations of CPU and GPU have the same tracking accuracy. Moreover, the GPU implementation has a higher processing speed and higher real time. And it even reached the requirement of High-speed tracking.DF tracking algorithm is a kind of tracking algorithm based on feature matching, its target feature is smoothed distribution field. This makes the tracking algorithm has some robustness. In the part of model training, one-dimensional and two-dimensional convolutions are both accelerated by GPU. The two-dimensional convolution is calculated by separable convolution. In target discrimination section, the computing discriminant value of candidate sample is accelerated by GPU. Tests show that, GPU can improve the speed of DF algorithm to 4.8 times of previous, reaches 45.8 frames per second.Struck tracking algorithm is a kind of tracking algorithm based on background learning. It uses the structured SVM to learn the difference between target and background, and get a discriminant function. The sample with the largest discriminant value is chosen as the target. In the part of model training, each iteration calculation module is accelerated by GPU. In target discrimination section, the computing discriminant value of all candidate samples is accelerated by GPU. Tests show that, GPU can improve the speed of DF algorithm to 4.3 times of previous, reaches 30.7 frames per second.KCF tracking algorithm is a kind of tracking algorithm based on background learning, it uses a circulation matrix to construct the training samples and has fast processing speed. This paper uses CUDA toolkit to accelerate the two-dimensional discrete Fourier transform. Tests show that, GPU can improve the speed of DF algorithm to 3.7 times of previous, reaches 488.2 frames per second.