Research On Visual Object Tracking Techniques Based On Correlation Filters

Visual object tracking is to estimate the movement of a target in an image sequence in this paper.As an important research field in computer vision,visual object tracking has various applications in human computer interaction and robot technology.In recent years,many excellent research results have been made in visual object tracking area.Correlation filter trackers(CFTs)have high tracking performance and computational efficiency,which attracts many researchers' last attention.However,the current research has the following problems:(1)it has insufficient adaptability to scale variations.At present,CFTs mainly adopt pyramid strategy to estimate scale.However,these trackers take the layer of the original target size as the center and then successively decrease the weight of the layers towards each side.When target scale changes,the fixed weight may weaken the real target scale response.And the discontinuity of the scale layer reduces estimation accuracy.(2)It has poor ability to deal with occlusion problem.Its periodic cycle limits the target search area,and the fixed learning rate method is easy to cause error accumulation and tracking failure when the target is occluded.(3)It is weakly robust under target deformation and rotation,for its template matching idea is highly dependent on the location of the target feature space.Therefore,focusing on the above problems,some research and exploration are carried out in this paper.The main work and results are as follows:1.A fine step scale estimation method is proposed.Based on pyramid strategy,this method carries out fine scale estimation in two steps.Firstly,the target is segmented according to its shape.During tracking,the trend of scale changes is judged by the relative position changes of the sub-graphs and the weights of the scale samples are offset.In this way,the real target scale response is enhanced and rough target scale estimation is achieved.Secondly,Newton method is employed to find the precise target scale.Experiments show that this method achieves more accurate scale estimation and effectively improves the tracking success rate.2.An anti-occlusion method based on sparse representation and online detection is proposed.Firstly,sparse representation model is used to distinguish the occlusion of the target.Secondly,According to the discriminant result,online support vector machine detection is exploited to achieve target relocation.Finally,in order to maintain the long-term stable memory of the target,the sparse representation model and support vector machine are conservatively updated according to the credibility of the current tracking results.Experiments show that this method can suppress occlusion interference effectively and gain higher distance precision and robustness on test sequences annotated with occlusion attribute.3.A tracking method based on HS histogram model and correlation filtering model fusion is proposed.The correlation filter tracking model is sensitive to deformation,while the color histogram tracking model has strong adaptability to deformation,but it is easy to be influenced by light change and complex background.Therefore,these two models are dynamically fused to achieve complementary advantages.Firstly,a background suppressed HS color histogram tracking model is proposed.The luminance component is separated to reduce the interference of illumination and a background weighted method is used to highlight object information.At the same time,to enhance the adaptability to background clusters and target deformation,a dynamic fusion strategy is proposed to adjust the fusion weights of the two tracking models dynamically according to the HS similarity between the target and the background.Experimental results demonstrate that this method performs well in regard of complex interference factors,such as deformation,illumination variation and background clusters,and has competitive tracking performance.