Research And Application Of Real-time Tracking Algorithm Based On Correlation Filters In Dynamic Scene

During the object tracking process,the tracked target is always in the status of motion.Tracking in a dynamic scene is prone to target occlusion,fast motion,motion blur,scale variation,deformation,in-plane rotation,out-of-plane rotation,out-of-view,illumination variation,low resolution,background clutters and son on.In such a complex dynamic scenario,model drift is prone to occur,will cause tracking failures,and in severe cases will cause permanent loss of follow-up.Based on the Correlation Filters(CF)algorithms of object tracking,I conduct in-depth research on Model Update and Features Fusion in object tracking.The article proposes three novel object tracking algorithms to enhance the robustness of the model,improve the problem of model drift,and ensure real-tracking.The main contributions of our works can be summarized as follows:(1)In dynamic scenes,the apparent characteristics of the tracked target are changeable,which easily pollutes the observation model and causes model drift.The article proposes a novel robust correlation filtering visual tracking algorithm with hand-drawn features to improve the drift of observation model in dynamic scenes.The characteristic response map will appear multi-peak when the tracked target is occluded,but not all peaks will affect the tracking effect.The main peak has the most impact on the secondary peak.The article uses a novel method to detect the primary and secondary peaks of the response map.The secondary peak of the response map has the greatest impact on model updates.Then,a new confidence function with good robustness in the correlation filter tracking algorithm is used to adaptively judge whether or not to update the model,which effectively reduces the probability of the model being polluted.(2)For the same problem of model update,most current object tracking models use a linear learning method with a fixed learning rate in model updating.Although the model can retain a lot of previous image frame information in this way,background information is easily learned into the model,which causes the model to be contaminated.The article proposes a new adaptive model updating method.To start with,a cosine similarity measure method is used to construct a confidence function to calculate the gap between the current frame retrieval target and the expected output target.In addition,using high-confidence frame images to construct a discrimination dataset which is updated by a voting mechanism.Finally,an adaptive learning rate is output according to the confidence value of the predicted image.The lower the confidence value is,the less effective information there is in the predicted image,so the model update learning rate is reduced,and vice versa.(3)At present in the field of object tracking,most tracking algorithms use dual or multiple features of the target instead of single target feature.However,simple linear addition is used in multi-feature fusion,which will greatly damage the performance of the tracking algorithm in dynamic scenes.Based on the Staple algorithm,a well-know object tracking algorithm,the article proposes a feature adaptive fusion tracking algorithm with the target probability model.To begin with,the target probability model is constructed based on the front-to-background ratio of the target,and then the averaged adaptive fusion coefficient or exponential adaptive fusion coefficient is selected based on relationship between the obtained probability value and the relevant threshold value,and the fusion coefficient value is calculated to achieve adaptive merge of multiple features.In order to show the impact of different merge coefficient thresholds on real-time tracking,in the article the GUI module of MATLAB is used to design an adaptive fusion complementary learning tracking system based on the probability model.In the article,the Online Object Tracking 2013(OTB-2013)dataset,Online Object Tracking 2015(OTB-2015)dataset and Temple Color 128(TC128)dataset are used to analyze the proposed algorithm experimentally.Comparing the performance of proposed trackers with state-of-the-art trackers of the same period,and quantitative and qualitative analysis underway.The experimental results indicate that the three trackers proposed in the article based on model updating and multi-feature fusion perform better,which have certain improvements in tracking accuracy and realize real-time tracking.