Research On Visual Object Tracking Algorithm Based On Correlation Filtering

Visual object tracking technology is one of the important branches in the field of computer vision.It is widely used in auxiliary medical diagnosis,military defense,virtual reality,visual navigation,etc.In recent years,the visual object tracking technology based on correlation filtering has attracted close attention because of its advantages of strong real-time tracking and low calculation cost.However,due to insufficient feature description ability and fixed object scale,traditional correlation filtering algorithms perform poorly in tracking scenarios such as illumination changes,object deformation and fast motion,which makes them difficult to apply to actual object tracking tasks.In view of the above problems and based on the kernel correlation filtering algorithm,this dissertation makes an in-depth analysis and improvement in the aspects of feature extraction method,object scale estimation,occlusion and redetection of lost objects and proposes a real-time robust correlation filtering object tracking algorithm.The main work and research are as follows.(1)Aiming at the single feature extraction method of kernelized correlation filtering algorithm has insufficient description ability,a weighted fusion multi-features object tracking strategy is proposed.First,two filters are trained independently based on the color and shape features of the object with the idea of ridge regression to determine the object position,then fusion the object position results of two filters in the response layer with the weight that determined by the approximate normalized peak sidelobe ratio.With the complementary characteristics of the two features,the tracking accuracy under the scenario of object deformation and illumination variation is improved.(2)Multi-scale estimation strategy is introduced.On the basis of considering the realtime performance of the algorithm while solving the problem of object scale variation during tracking,this dissertation constructs an object sample set of different scales centered on the object position determined by the current frame,and then the sample set is used as the input to train a one-dimensional scale filter which is used to determine the best scale of the object in real time.This strategy improves the tracking accuracy of the algorithm in the scene where the object scale changes.(3)Aiming at the problem of object loss and drift caused by occlusion,rapid object movement and blurring in the algorithm tracking process,an object resampling detection update strategy is proposed.When the algorithm fails to track in the current frame,the object position in the previous frame is used as a reference to construct a resampling detection area,and the object is rechecked in this area to determine the position of the object.Finally,according to the tracking confidence of the recheck results,the algorithm determines whether the template of the position filter and scale filter is updated,which reduces the possibility of introducing background interference and improves the robustness of the algorithm.In this dissertation,the effectiveness of the proposed algorithm is verified on a benchmark dataset containing 85 sets of video sequences with different challenge attributes.The test results show that the algorithm in this dissertation can well adapt to the situation of object deformation,scale change,occlusion and fast motion in the actual tracking scene,and fully meet the real-time requirements of the algorithm,and has strong practicality.In addition,compared with the traditional correlation filtering algorithm,the algorithm in this dissertation has obtained higher tracking accuracy and success rate,showing a more superior performance.