Research On The Weighted Multi-feature Fusion Of Appearance Model In Object Tracking

Visual tracking is one of the important research topics in the field of computer vision and widely applied in intelligent monitoring,human-computer interaction,and so on.At the same time,due to the challenges such as illumination,occlusion,scale,rotation and other factors,there is still much room for development in current tracking research.Correlation filter-based trackers have made excellent performance due to their efficiency in recent years.However,the description of object appearance with computer to improve the robustness of the model is always the most important premise for the algorithm design.Aiming at the problems of weak robustness and inaccurate target location of existing tracking algorithms,this paper accomplishes the following works with correlation filtering algorithms:Firstly,on the description of the target appearance model,a new weighted multi-feature fusion model HCG is proposed to model the target in this paper.With the correlation filtering framework,this paper introduces HOG?color and original pixels features and analyses their limitations of each single feature and shortcomings of existing feature combination methods in describing objects.Combining feature dimension information,this paper designs the proportional relationship between features and constructs weighted vector to fuse the features.Finally,experiments show that the model can compensate for the shortcomings of single feature,and avoid the existing feature fusion method too biased to a single feature,which makes other features ineffective.Secondly,aiming at the problem of inaccurate target location in target detection,this paper proposes a target relocation method for peak reliability evaluation.The maximum peak value and average peak correlation energy are used to judge whether the current target location is accurate or not.For the accurate target location,the highest peak corresponds to the target center,and the feature model HCG is updated in this paper.Otherwise,this paper chooses multi-peak with response graph to determine the suspicious location of the target.Then taking multi-peak as the center,comparing the similarity between the features of surrounding image blocks and the HCG model stored in the tracker.Finally,the reliability factor is introduced by the similarity to evaluate the reliability of the peak value,and the peak value with high reliability is taken as the target position in the new frame.Finally,based on the above work,a tracking algorithm with weighted multi-feature fusion appearance model called WMFF(weighted multi-feature fusion model)is proposed in this paper.The tracker combines the construction of the weighted multi-feature fusion target model and the target relocation based on the reliability evaluation of peak value.This paper gives the complete algorithm steps and flow chart.In the open datasets,the proposed algorithm is compared with the current popular algorithms qualitatively and quantitatively.The experimental results show that the proposed WMFF model not only complement each other's features,but also eliminate the interference of similar objects and occluded scenes.