Research On Uncertain Motion Object Tracking Method Based On Region Proposal

Visual tracking is a fundamental problem in computer vision,and it has a wide range of applications in areas such as video surveillance,automatic vehicle navigation,and humancomputer interaction et al.Although the study of visual tracking technology has made significant progress.However,in practical applications,there are often numerous challenges to visual tracking due to the fast motion,abrupt motion,cross-vision motion and other uncertion motion scenarios.Therefore,how to design an accurate,robust and generalizable object tracking algorithm is a key problem to be solved.In this dissertation,the region proposal method is introduced into the traditional target tracking framework to achieve long-term tracking under uncertion motion scenarios.The main research work and innovations are described as follows:(1)An object tracking method based on semantics estimation region proposal and deep correlation filter is proposed.Firstly,a semantics region proposals generation strategy is presented,including category-level semantics proposals,one-object-level semantics estimation and semantics-contextual proposals generation,to obtain a few of high-quality object-oriented proposals which can cover the object state under uncertain motion scenarios.Secondly,combining the globally sparse semantics region proposals prediction and correlation filter prediction,a hybrid semantics tracking algorithm is proposed,which obtains the coarse object location by the decision of multiple response maps.Finally,independent correlation filters are learned and trained to estimate the scale of the target for a higher tracking accuracy.Extensive experiments on tracking benchmarks demonstrate the proposed method achieves state-of-theart performance.(2)An object tracking method based on light regression memory network and multiperspective region proposal is proposed.Firstly,a highly accurate multi-perspective object special proposal method is proposed to recover tracking when tracking failure happened.Commonsense information is used as a priori knowledge to refine region proposals.Meanwhile,semantic target-aware estimation and spatial structure estimation are designed to rank and choose region proposals from two complementary viewpoints seperately.By a comprehensive decision,the final proposals can fully cover the target's states even under abrupt motion.Secondly,the light regression memory network is learned on a single-convolutional layer network by convolution linear regression.The change of target's appearance can be reinforced by fine-tuning a fixed number of network parameters are online with reliable target frame information.Without increasing size and complex update strategy,the light regression memory network can determine whether tracking fails.Finally,the proposed memory network and region proposals are integrated into Siamese-based tracking framework to achieve accurate uncertain motion tracking.Numerous experiments on multiple tracking benchmarks prove that the proposed tracker achieves excellent performance.