The Research Of Visual Object Tracking Algorithm Based On GPU Acceleration

Visual target tracking is one of the hot research topics in the field of computer vision.It is widely used in video monitoring,human-computer interaction,intelligent transportation,military strike and other fields,and has extremely important practical significance.In recent years,great progress and achievements has been made in target tracking,however,the actual tracking scene is often very complex with a variety of interference factors,considering both tracking accuracy and tracking speed is still a difficult problem to be solved urgently.In this paper,we optimize the DSST algorithm to get better performance,and use the GPU to accelerate the improved algorithm to achieve real-time tracking.The main research contents of this paper are as follows:1.We optimize the DSST algorithm,design the redetection module,and propose a long-term target tracking algorithm DSST-RD.The DSST algorithm is a tracking algorithm based on correlation filtering,which divides target tracking into two problems: target position translation and target scale transformation for the first time.This algorithm can solve the problem of target scale transformation well,but it can't be relocated when the target is seriously occluded or the target disappears from the current frame.Therefore,this paper designs a redetection module based on outlier detection and correlation filter classifier,the search strategy of local search first followed by global search is adopted here,and the model library composed of multiple correlation filter classifier models is used to realize the target relocation.During the global search,we use a variance classifier to reduce the candidate target area.Then,on the basis of the DSST algorithm and combined with the redetection model,we designed a long-term target tracking algorithm model that can deal with the problem of target loss.Finally,we test the tracking performance of the improved algorithm by using the standard test set OTB100.The results show that the improved algorithm has better tracking accuracy and robustness,as well as good tracking performance for target occlusion and redetection.2.We use CUDA technology of GPU platform to optimize and accelerate the DSST-RD algorithm.DSST-RD tracking algorithm has good tracking performance,but its tracking speed is slow,and it cannot meet the real-time requirements.Therefore,we use CPU+GPU heterogeneous architecture to speed it up in parallel.Firstly,we analyze the time consuming and dependencies between modules in the algorithm,and determine to adopt data parallelism to design the CUDA program.Then,we use CUDA heterogeneous parallel programming model to conduct parallel design for the FHOG feature extraction,kernel correlation calculation,fast Fourier transform,fast detection and model update module,and carefully design the code from two aspects of thread organization and memory model selection.Finally,we optimize the overall memory,allocate and release memory uniformly,and reduce the data copy operation between the device and the host,to reduce data transmission delay and improve tracking speed.The test results in the standard data set show that the GPU-accelerated algorithm can maintain the tracking success rate of the original algorithm and achieve real-time target tracking.