Research On Object Tracking Algorithm Based On Non-symmetry And Anti-packing Model

Target tracking is one of the basic core technologies of computer vision.It is closely related to our daily life and production,and has great application value in many fields such as civilian and military.However,since it is easily affected by the target's intrinsic factors,such as nonrigid deformation,rotation,and external environmental factors such as illumination change,occlusion and scale variation,the development of target tracking still faces great challenges.It is still a very difficult task to design a target tracking algorithm that can balance the accuracy,robustness,and anti-jamming of the tracking and have good real-time performance.The Nonsymmetry and Anti-packing Model is a very advanced image representation method,which can highlight the structural characteristics of the image with fewer blocks.In this paper,based on the Non-symmetry and Anti-packing Model,the target tracking algorithm is studied in two aspects:(1)An object tracking algorithm based on the Non-symmetry and Anti-packing Model and the local density estimation is studied.The main contributions of this algorithm are summarized as follows: Firstly,The image segmentation algorithm based on the Non-symmetry and Antipacking Model is used to divide the target image into homogeneous blocks to construct the target segmentation template,and then it is combined with the fast local density estimation to establish the target description model.The fast local density estimation makes our algorithm achieves excellent real-time performance.Secondly,a model updating technique using local adaptive learning and negative feedback adjustment is proposed to alleviate the occulusion interference and model drift.Thirdly,we improve the median flow tracking for estimating the scale variation in order to reduce the interference caused by the scale change.(2)A compressive tracking algorithm based on the Non-symmetry and Anti-packing Model and structured SVM is studied.The algorithm combines the advantages of the compressive tracking algorithm with high real-time performance and strong discriminative ability of the structured SVM.Its main contributions are summarized as follows: Firstly,the image representation method based on the Non-symmetry and Anti-packing Model is introduced in the process of extracting compressed features by the compressive tracking algorithm to enhance the ability of the compressed features to express the target.Secondly,a coarse-to-fine multi-peak target detection mechanism is proposed to reduce the effects of background clutter and similar target interference.Thirdly,a scale estimation method based on the scale pool model is proposed to eliminate the interference of scale changes.Finally,a high-confidence sparse update strategy is proposed,which significantly reduces the classifier learning noise samples,avoids model drift,and improves tracking speed.The experimental results demonstrate that the tracking algorithm based on the Nonsymmetry and Anti-packing Model and local density estimation is superior to most state-of-theart compared algorithms,and achieves a tracking speed of 87.33 FPS on the OTB100 data set,and also has obvious anti-interference ability in terms of occlusion,deformation and scale change.The compressive tracking algorithm based on the Non-symmetry and Anti-packing Model and the structured SVM performs superiorly against most state-of-the-art compared algorithms,and achieves a tracking speed of 41.57 FPS on the OTB100 data set,which basically meets the needs of real-time performance,and has obvious anti-interference ability in terms of background clutter,target deformation,occlusion,scale change,motion blur,etc.