Adaptive Tracking Algorithm Based On Correlation Filtering

As one of the hot spots in the field of computer vision,target tracking has been widely used in many fields,such as automatic driving,national defense security,smart home,human-computer interaction,etc.At the technical level,target tracking involves pattern recognition and image processing,which is a research direction of wide field and multi application.In recent decades,with the rapid development of the field of target tracking,a large number of excellent algorithms have emerged and achieved good results to a certain extent.However,in practical application,the tracking environment is changeable and complex,such as fast movement,illumination change,occlusion,background interference,etc.In the face of these difficulties,how to achieve robust and real-time tracking is still a difficult but meaningful task.This paper analyzes the advantages and disadvantages of different types of tracking algorithm,it uses the correlation filter tracking algorithm as the research framework,the problems such as occlusion,deformation,rapid movement and background interference in the target tracking process are studied in depth and an improvement scheme is proposed.The advantages of the improved algorithms are proved by experiments,which are mainly divided into three aspects:(1)In view of the fact that the correlation filtering algorithm is easy to cause tracking failure in complex situations such as target occlusion and fast motion,and is difficult to be applied to long-term tracking,this paper proposes an adaptive correlation filtering algorithm suitable for long-term tracking.Firstly,the HOG feature,CN feature and gray feature are connected in series to enhance the feature discrimination.Then,the Edgeboxes are used to generate detection suggestions,and the optimal candidate box is found to realize the adaptive scale and aspect ratio of the tracker.In order to avoid the damage of the template,a new adaptive update rate is formed by combining the target's speed and the number of edge groups,and the high confidence is judged.The motion estimation method is used to further correct the scale of each frame.Finally,in the case of tracking failure,the SVM re-detector is used to restore the target position.(2)In view of the fact that the single feature of correlation filtering algorithm leads to the lack of discrimination power of the target representation model,this paper proposes an adaptive cosine window multi-feature correlation filtering algorithm.Firstly,the bayesian classifier based on color histogram is adaptively fused with the traditional cosine window to further suppress the influence of object deformation and background clutter.Then,in the training process,the hierarchical depth feature and manual feature are extracted respectively,and the improved cosine window is added to the feature map.The response was fused in a fixed proportion.Then the scale pyramid model is used to obtain the optimal scale.Finally,when the model is updated,the average value of the confidence score calculated by three different feature graphs is calculated to further improve the anti-occlusion performance of the algorithm.(3)In view of the fact that the real training samples of correlation filtering algorithm areinsufficient,and the detection range is limited.This paper proposes an adaptive correlation filtering algorithm based on Kalman filter.Firstly,PCA algorithm is used to reduce the dimensionality of the fusion features,reduce the computation amount of the algorithm and improve the running speed.Then the binary mask matrix is added to the loop sample clipping to increase the number of real samples to train the correlation filter and alleviate the boundary effect caused by the loop matrix.Then the Kalman filter is introduced to estimate the position of the tracking target and adjust the detection range.Finally,a sparse template update strategy is adopted to accelerate the algorithm.In this paper,the OTB-100 and UAV20 L datasets are compared with a variety of mainstream algorithms,and the performance of the improved algorithm is analyzed quantitatively and qualitatively.Experimental results fully demonstrate the effectiveness of the three algorithms proposed in this paper.Aiming at the adaptive correlation filtering algorithm based on Kalman filter,due to its fast operation speed,Matlab GUI technology is used to call the external camera of the computer in real time to verify the practicability of the algorithm.