Research On Generic Target Tracking Technology Method Based On Apparent Learning

Object tracking in video sequences is one of the key techniques in the area of computer vision and artificial intelligence,which can help computers to locate the 2-dimensional position of target in continuous frames and provide further module with useful information.Benefiting from this technology,many applications have come true.As the development of machine learning and other emerging technologies,object tracking is of more and more diversity and engineering.Expressly,robust appearance model and ingenious strategy have decisive effects on tracking performance,which has been confirmed by the previous work.The research of generic object tracking methods based on online learning is a key and challenging issue of this technology.In this paper,plenty of work is devoted to studying the relevant theories and designing the framework.The main contributions are listed below:(1)the fundamental problem of object tracking is divided into several sub-problems,based on which the related research and overview are carried on separately.Besides,a deep analysis of the current research status is provided,and several typical algorithms,e.g.,particle filter,support vector machine and correlation filter,are also discussed.(2)a new support vector machine and particle filter-based tracker using discriminative model and information interaction strategy is proposed.Specifically,we focus on establishing a robust discriminative tracking model with joint color and texture histogram-based feature vectors via the linear support vector machine method to achieve satisfactory performances in challenging scenes.Intended to fit the particle filter,the outputs of the model are mapped into probabilities with logistic regression.Then,the strategy of information interaction aiming to replace the complex filtering with the high-confident correlation result in adjacent frames is added.To avoid the problem of sample accumulation,we sum up the optimal model using an instance set with the predefined budget,and update the classifier on both the previous set and the updated data from the tracking results every few frames.(3)in light of the drift and occlusion problems existing in the typical correlation filter tracking algorithms,we propose a collaborative tracking strategy based on bag-of-visual-word model and correlation filter.Its main idea is to help the appearance models with two different visual cues to complement with each other,and to make anomaly monitoring as well as tracking restoration in the basic tracking process come true.Specifically,we introduce a robust subpixelbased bag-of-visual-word model to the original method and recognize the abnormal tracking result with low confidence by calculating the foreground confidence in the object region.Under the abnormal circumstances,the tracking process can be restored using a proposed strategy that combines visual attention and multi-filter templates detection,thereby accomplishing constructing a robust and reliable tracking system.(4)we program and run the codes of the proposed algorithms on the software platform of Matlab,and test their performances on the public databases.The robustness and the effectiveness of our methods are verified by result assessments and comparative analyses.