Multi-modal Data Based Fusion Tracking Algorithm Research

The rapid development of computer science and technology has led to the innovation in the field of computer vision.The increasing demand for machine learning and artificial intelligence has made visual object tracking a hot topic of current research.Visual object tracking will profoundly affect the development of driverless car,security,human-computer interaction,navigation and guidance in the future.After decades of development,current object tracking algorithms still face challenging interference factors from external environment and the object itself,such as background clutter,occlusion,low illumination,scale variation,deformation,motion blurring and fast motion,which seriously restrict the development of object tracking.In this paper,by studying the complementary characteristics of data from different modalities and the advantages and disadvantages of different tracking methods,a multi-modal fusion tracking algorithm based on the "Tracking-by-detection" framework is proposed.The algorithm adopts the global and local features of the target from infrared and visible images,and can deal with multiple complex interferences in the current object tracking field.Firstly,two tracking modules are designed,one is based on statistical model(HIST module)and the other is based on correlation filter(CFT module).In HIST module,RGB color histogram with global statistical characteristics is used as tracking feature,and a target/background discriminant operator is designed according to Bayesian criterion to distinguish object and interferences.It is a hybrid tracking model of generative and discriminative tracking strategies.In order to improve HIST module according to “Tracking-by-detection” framework,integral image is introduced to make the fusion between HIST and CFT module possible.The CFT module which is a discriminative tracking model,adopts multiple features(HOG,CN,image intensity)during tracking according to the principle of KCF.In this module,a denoising fusion rule is designed to fuse the response maps obtained from various features.And through detailed theoretical analysis and derivation,an improved CFT module is proposed according to “Tracking-by-detection” framework.Secondly,according to KL divergence,a reliability measurement is proposed to measure the reliability of the tracking results of the above two modules.Based on the results of the measurement,this paper also proposes a decision-level adaptive fusion strategy,which can get the final tracking results by adaptively fusing the tracking results of the two modules.Finally,by using the standard VOT evaluation benchmark and 48 pairs of visible-infrared video sequences from RGB-T234 dataset,this paper designs efficient experiments for evaluation.There are 10 advanced mainstream tracking algorithms based on single-modal or multi-modal data are compared with the proposed method.The final analysis results are given from three aspects: quantitative analysis,attribute-based analysis and qualitative analysis.Experimental results show that the proposed visual object fusion tracking algorithm based on multi-modal data is accurate and robust even under complex interference factors such as background clutter,occlusion,low illumination,deformation and scale variation,motion blurring and so on.It proves that the proposed algorithm has important value in theoretical innovation and real-world applications.