Research On Object Tracking Algorithm For Scale Variation And Occlusion Problem

Visual object tracking is a significant research topic in the field of computer vision,which widely used in security surveillance,autonomous vehicles,human-machine interaction,and military reconnaissance.To decrease the tracking failure rate and increase the accuracy of tracking,lots of researchers have deep study in this topic,which has enabled object tracking to achieve rapid development in recent years.Correlation filterbased tracking algorithms have become one of the important methods for object tracking because of their strong discriminative ability and fast calculation speed.However,limited by some interferences in the complex environment,the accuracy of correlation filters tracking methods needs to be further improved,especially in scale variation,occlusion,similar background interference,and fast motion.Therefore,this thesis focuses on the problems of scale variation and occlusion,and proposes improvements from the aspects of scale estimation,object state,and real-time model update.The main research contents of this thesis are as follows:1.Concerning the problem of insufficient object scale tracking accuracy in correlation filters-based tracking algorithms,this thesis proposes a scale estimation network that uses a large-scale data set offline training.First,this thesis analyzes the problems for correlation filter-based tracking algorithms using scale filters.Then,an improved scale estimation network is proposed,which uses a deep feature network to extract image features and calculates the value of intersection over union for candidate bounding box through the pseudo-siamese network.The object scale is obtained by optimizing the result of intersection over union value.Finally,this thesis proposes an object tracking algorithm combines the scale estimation network and a classification filter,which discriminates the object and background.Experiments are performed on OTB-100 and VOT object tracking datasets,and results show that the proposed tracking algorithm has improved the tracking accuracy.2.Regarding the problem of high tracking loss rate for correlation filter-based tracking algorithms in occlusion,this thesis proposes a discriminated formula of object occlusion and an adaptive module update method.First,this thesis analyzes and summarizes reasons that correlation filter-based tracking algorithm fails to track in occlusion.Then,this thesis proposes a discriminated formula of object occlusion to get an object occlusion state.Object occlusion state is judged by comparing the values of the object's continuous multi-frame features correlative value and an occlusion determination threshold obtained by experiments.This thesis considers that tracking object is occluded when correlative value is less than the occlusion determination threshold.Meanwhile,this thesis proposes an adaptive filter update method to improve robustness by controlling the updating rate of the filter and reducing the weight value of the occlusion object.Experiments results show that the adaptive filter update method to improve the tracking accuracy effect is not obvious,and the model update rate is not the main factor that reduces the ability of correlation filters' discriminate.