Research Of Algorithm On Human Action Recognition Based On CUDA

With the development of computer technology and the coming of the intelligent era,the intelligent recognition technology of the action of the human in the video surveillance system has attracted more and more attention,and the importance of the human action recognition algorithm is becoming more and more important.One of the most important steps in the human action recognition algorithm is to extract the movement trace of the body parts.The accuracy of the results of the action recognition is closely related to the accuracy of the extracted trajectories.Many methods such as HOG,SIFT,optical flow method have been proposed for feature extraction.However,in many scenarios,it is still difficult to achieve accurate and robust extraction of the trajectory of the human by these methods.It is difficult to practice the human action recognition technology.In this paper,the Exclusive Block Matching(EBM)method is used to extract the moving track of the characters,and the Context Histogram of Trajectory(CHOT)is used to describe the trajectory characteristics.The results show that the algorithm has good performance in trajectory extraction and feature description.In order to improve the recognition accuracy,this design feature classification method of action figures based on bag of words model,the action of as synthesized by a series of small movements of the independent model to describe movement.EBM method is only applicable to the static background,and now in the robot vision,intelligent transportation systems,video surveillance systems and other fields have a lot of dynamic background of the human in the video to identify the needs of the action.However,due to the impact of vulnerable to illumination and viewpoint changes,changes in the size and noise of some phenomenon,such as relying only on the HSV histogram,edge histogram,HOG histogram and SIFT features of visual features such as information is difficult to achieve accurate tracking of the human.In order to achieve robust tracking,we extend the EBM method,put forward a consideration of visual feature information and structure information of the template matching method,capable of specific human in the video to achieve robust tracking,and can determine the relationship between the human in the video part of each part of the template and the people.EBM method and template matching method have very good accuracy,but the algorithm complexity is high,can not meet the requirements of real-time.In the calculation process of the algorithm has a lot of data for the same operation of the task,the task is very suitable for parallel computing.And GPU is a kind of processor which is dedicated to the highly parallel computing.In this paper,the characteristics of GPU based on CUDA is proposed for heterogeneous parallel computing optimization scheme based on analysis of the algorithm is suitable for parallel computing,and through the CUDA C programming language,the parallel computing task performed by GPU.The whole algorithm through the CPU and GPU collaborative computing,greatly improved the speed of calculation.Finally,after experimental verification,a large amount of data and draw the conclusion that human action recognition compared to the traditional algorithm,EBM+CHOT feature extraction method combining description methods of word classification bag model is applied to human action recognition,we obtain accurate recognition results.The template matching algorithm based on visual feature constraints and structural constraints can accurately track the human and obtain the corresponding relationship between the object and the template.CUDA based parallel optimization program can effectively improve the calculation speed of the algorithm,the maximum speed up to 13 times,there is a great application value.