The Research Of Visual Object Tracking Algorithm And Its Optimization With CUDA Acceleration

Visual object tracking is a classical computer vision problem with many extensive applications in areas such as intelligent robotics,video surveillance and automatic drive.It imposes major challenges,such as appearance changes,occlusions and background clutter.Over the past few decades,the tracking model complexity and compute needs dramatically increase with the tracking algorithm becoming more and more accurate,and its real-time performance is poor.CUDA(Compute Unified Device Architecture),a general-purpose parallel computing platform and programming model,makes real-time tracking into a reality in visual object tracking.In this thesis,we focus on the study of tracking based on keypoints,correlation filter tracking Beside,CUDA heterogeneous programming technology has been employed to optimize the tracking algorithm to meet the real-time requirements.Therefore,the research of this thesis has certain theoretical and practical reference.The primary work of this thesis as follows:1.We propose an enhanced CMT(Clustering of the Static-Adaptive Correspondences for Deformable Object Tracking)tracking algorithm with CUDA acceleration method.CMT based on keypoints and geometric dissimilarity measure is robust deformable object tacking.Its optical flow tracker is error-prone when active points run outside the scope of the target and leads to tracking failure.In order to solve this problem,we make use of training the HOG features of samples to construct a strong SVM model,which helps filter wrong active points and relocate the target position to improve tracking precision.In order to further improve the speed of enhanced CMT algorithm,a kind of CUDA parallel optimization method is proposed in this thesis.Enhanced CMT parallelized by CUDA heterogeneous parallel programming significantly improves computation efficiency.Ultimately,experimental results demonstrate that on the promise of higher tracking accuracy,our proposal can achieve real-time tracking.2.We propose an adaptive long-term visual tracking algorithm with CUDA parallelization.Based on CMT tracking algorithm,we build static and dynamic models based on keypoints for long-term tracking to provide long-term and reliable information.Meanwhile,a multi-scale correlation filter pyramid with powerful HOG features and color attributes to realize better localization and search for the optimal scale exhaustively for short-term tracking.Besides,we propose an adaptive model update method for these two complementary parts to further boost the overall tracking performance.In order to re-detect the object,we extract and compress object HOG features with a compressed sensing matrix to build a sparse support vector machine model.In order to improve the real-time performance of the algorithm,CUDA has been employed to accelerate proposed method to achieve real-time tracking.Extensive experimental results illustrate outstanding performance in terms of target scale change,occlusion and re-detection compared with the state-of-the-art algorithms,and it can achieve real-time and long-term tracking.