Robust Object Tracking Based On Ke Rnelized Correlation Filter And Its Application In Indentifying The Actions Of Cows

Object tracking is one of the hot topic in computer vision.It is widely used in civil and military fields.However,robust object tracking is difficult due to the deformation,rotation,scaling from target object and the complex disturbs from external environme nt such as illumination changes and object occlusion.In recent years,object tracking methods based on kernel correlation filter have achieved great improvement.The basic idea for this kind of tracking method is that it density samples the target and its surrounding background,and converts the solution of the classifier into the Fourier frequency domain to achieve object tracking.Therefore,this kind of tracking met hod can extremely utilize the features of target object and its surrounding background,which finally leads to the accurate and robust tracking results.Based on the kernel correlation filtering,this paper introduces the saliency detection of object and the patch tracking to construct two object tracking methods.Then,the proposed object tracking method is used to identify the actions of cows to guide the healthy farming of cows.The main contributions of this paper are as follows:(1)An object tracking method is proposed by defining the combined kernelized correlation filter.This combined kernel correlation filter is composed via two kernel correlation filters and a combination processor.Firstly,in the color domain,using the kernel correlation filter,a intrinsical target appearance model based on the target object region and a target background appearance model based on the enlarged target object region are constructed intrinsical.At the same time,a target saliency appearance is defined based on the significant values about the target object and the cornel correlation filter.Here,the image saliency method proposed based on the minimum spanning tree is used to compute the significant values.Then,after the two kernel correlation filters are utilized in the color domain and saliency domain,a combination processor is defined to combining the two filters to construct a robust object tracking method.In detail,the main processes of this proposed method are as follows: predict the target position by the kernel correlation filter based on the appearance model of the target background in the color domain,using the combination processor to validate the predicted target position,utilize the kernel correlation filter based on the target saliency appearance model when the validation is poor to overcome the disadvantages of kernel correlation on object occlusion and illumination changes and so on.Many experimental results show that the proposed object tracking method has good effect in dealing with the tracking challenges such as occlusion,illumination change and fast motion,and it is superior to the existing tracking algorithm on the above challenges,which shows a certain superiority.(2)A target tracking algorithm based on identifying the optimal target patches is proposed.Firstly,the appearance models of many target patches are constructed and tracked by the kernel correlation filter.Secondly,the mixed Gaussian model is used to replace the traditional Gaussian model to optimize the training of the classifier which is the key point of the kernel correlation filter.With the optimized kernel correlation filter,the proposed method can accurately track the discrete target patches.Thirdly,using the Hough voting and the tracking results of target patches,the proposed method computes the target position of tracking object.Finally,the proposed object tracking method updates the appearance model by separately updating the normal patches and the abnormal patches,which can greatly adapt to the changes of target object and its surrounding background.A large number of verification experiments show that the algorithm has certain superiority in position accuracy and tracking success rate,and has a good effect in dealing with the challenges of background clutter,light change and size change.At the end of this paper,the proposed method is applied to indentify the actions of cows.The health status of dairy cows is closely related to the behavior of daily life and the health status of dairy cows has a direct impact on the milk production,milk and other milk indicators.Through the object tracking of a cow,the movement of the cow can be identified by the reference measure such as the speed of the target displacement,the relative distance of the displacement and the size change of the target,and the trajectory can be analyzed by analyzing the trajectory.By analyzing the movement direction,movement trend and movement scope of the cow,people can monitor of the behavior of the cow and identify its actions such as the simulating action and go home action.The actions of cows can guide farmers to achieve the precision and healthy breading cows,which finally leads to improving the quality and quantity of cows.The experimental results show that the proposed algorithm performs very effectively and efficiently in identifying the actions of cows.