Development And Optimization Of Target Tracking Algorithm Based On GPU

Target tracking is one of the key issues that are concerned in the field of intelligent video analysis. It has an important role in many application fields, such as security monitoring, traffic monitoring, video synopsis, human-computer interaction, the military and vehicle navigation. However, Target tracking with both robustness and computation efficiency is still a challenge in face of complex and varied natural scene.This thesis analyzes state of the art target tracking algorithm, then it developed the algorithm based on RPT target tracking algorithm. Building patches in the basis of normal distribution probability model, it extracted the 31-D FHOG feature from the gray image of target patches. Using the characteristics of cyclic matrix and Fourier transform, it transforms the calculation to the Fourier domain. Then, it locates target patches through kernel correlation and closed-form model, and it traines and updates the model. It achieves the goal of target tracking by using all the tracking information and motion information of target patches to vote and screen. In order to improve the computation efficiency of the algorithm, it analyzes the exection time and parallelism of each module in the algorithm implementation. In view of FHOG feature extraction, kernel correlation calculation, model training and update and other models that are with more execution time, it designs optimization methods with parallel computation using CUDA on GPU, and it implements the paralleled optimization of target tracking algorithm based on GPU.Experiments show that the design of GPU parallel computing optimization methods in this thesis improves the executiong speed of tracking at the same time maintaining the accuracy. Compared with the CPU serial implementation, the acceleration of the GPU optimization implementation is up to 6.66times, which has good engineering application value.