Research Of Object Tracking Algorithm Based On Part-based Correlation Filter Using Parallel Computation

With the development of computer vision and artificial intelligence,using visual analysis algorithms to simulate the human eye and brain to realize the understanding of images and videos has become the research hotspot in recent years.Visual object tracking is one of the most important topics in the field of visual analysis such as artificial intelligence.It has extremely important applications in areas such as unmanned driving,traffic detection,security monitoring,human-computer interaction,medical applications,navigation,and military.But when faced with changes in appearance such as rapid motion,motion blur,shape change,target rotation,scale change,and out of view,and environmental changes including background clutter,light changes,low resolution,and partial occlusion,visual object tracking is still an open challenging task.This thesis studies the advantages and disadvantages of the tracking algorithms based on correlation filter,and proposes a tracking algorithm framework based on multi-part strategy in the Bayesian framework.This algorithm mainly solves the most challenging interferences like occlusions,deformation and scale variation in visual object tracking.The core components in this framework proposed in this thesis include the appearance model,motion model and feature model of the target.In this thesis,the tracking algorithm based on the correlation filter is used to address the appearance model problem,exploit the temporal relationship among these parts to calculate the reliability of each part,and use the relationships of parts in one single frame to define the contribution that each part made for the entire tracking result.In addition,an adaptive update scheme for each model and parameter and an adaptive scale estimation method for the target are proposed.Moreover,in order to reduce the computational complexity of the multi-part tracking algorithm,this thesis analyzes the execution time ratios of various components of traditional correlation filter based algorithms.And in order to realize real-time visual object tracking,the CUDA software technology on the GPU hardware platform is used on the components that can be computed parallelly,including feature extraction and kernel function matrix calculation.Finally,CUDA's stream processing capability is used to achieve the task level parallelization of the whole tracking during one tracking time,which greatly improves the performance of the proposed algorithm in terms of real-time tracking.Finally,a comprehensive evaluation of the proposed algorithm is carried out on 100 video sequences containing various types of challenging interference.The proposed algorithm is compared with the most mainstream 10 algorithms on quantitative analysis,qualitative analysis,and attribute-based analysis.The experimental results show that the parallel algorithm framework designed in this thesis not only can achieve stable and accurate tracking under the interferences including occlusion,deformation,and scale variation,but also outperforms traditional part-based algorithms in terms of real-time tracking.The proposed algorithm has important value both in theoretical innovation and in real-world engineering application.