Visual Feature Adaptive Selection And Fusion Method For Robust Tracking

As a hot topic in the field of computer vision,moving object tracking in video image is of great significance.However,object tracking under complex circumstances is still a challenging task because of background interference,obstacle occlusion,object deformation etc.Robustly detecting,locating and analyzing a target under complex circumstances through single-feature representation are difficult tasks.An effective method to address this problem is adaptive fusion of multiple features in representing targets.The process of adaptively fusing different features is the key to robust object tracking.Based on the selection and fusion of multiple features,this paper analyzes the problem of object tracking under complex circumstances.The main work of this paper is as follows:Firstly,the research status and development trend of multi-feature moving object tracking are introduced.And the relevant theoretical basis of multi-feature moving object tracking is expounded and analyzed.In the second place,through the analysis of traditional feature selection methods,a feature selection method which is based on features' stability and contrast evaluation is presented.In this feature selection method,variance is used to measure stability and contrast ability is calculated by log likelihood ratio.Then,the importance of each feature is evaluated by stability index S and the contrast index C.Thus,features which are good for the subsequent object tracking are selected.The correctness and effectiveness of the proposed feature selection method are proved by a large number of convincing experiments.This method has also laid a good foundation for the subsequent multi-feature adaptive fusion tracking.Finally,based on Mean Shift tracking framework,this paper proposes a multi-feature joint descriptor(MFJD)and the distance between joint histograms is used to measure the similarity between a target and its candidate patches.Color and Histograms of Oriented Gradients(HOG)features are fused to represent the tracked object and a multi-feature adaptive fusion strategy on the basis of mean shift framework is further proposed.This paper also proposes a self-adaptive multi-feature fusion strategy that can adaptively adjust the joint weight of the fused features based on their stability and contrast measure scores.Meanwhile,lots of experiments which are from the qualitative,quantitative and adaptive strategy evaluation demonstrate that the proposed MFJD method is effective in dealing object tracking under complex scenes,such as occlusion,posture variation,background blur and object size change.Furthermore,by comparing with large numbers of contrast algorithms the results show that the proposed adaptive multi-feature fusion tracking method MFJD are better than other advanced algorithms in tracking accuracy and robustness.