Adaptive Scale Object Tracking With Kernelized Correlation Filters

In computer vision,target tracking has important roles in many fields,such as imaging guidance,video surveillance,motion analysis,and human-computer interaction.In essence,target tracking is to obtain information such as the position of the target of interest in the image by processing non-stationary,time-varying targets and background image streams.After many years of research,many tracking methods have been proposed,but the tracking of light changes,occlusion,poses,and scale changes is still difficult.Because most existing detection-based trackers do not solve the scaling problem,this paper designs a scale estimation strategy based on the traditional detection-based target tracking framework,and presents an adaptive scaling target based on nuclear correlation filters.Tracking algorithm.The algorithm uses a kernel function to solve a regularized least squares classifier to obtain a nuclear correlation filter,and acquires image samples through scale-invariant feature transformation.Through on-line learning of the nuclear correlation filter,the target position and scale are detected,and online.Update the nuclear correlation filter.SIFT is a local feature descriptor proposed by David Lowe in 1999 and was further developed and improved in 2004.SIFT feature matching algorithm can handle the matching problem in the case of translation,rotation,affine transformation between two images,and has a strong ability to match.In invariant comparison experiments with ten local descriptors including SIFT operators,SIFT and its extended algorithm have been proved to have the strongest robustness in the same class of descriptors.On the other hand,a nuclear correlation filter is used,in which the circulant matrix diagonalization has a magical effect,which greatly reduces the amount of computation.The correlation filtering method trains a correlation filter based on the information of the current frame and the information of the previous frame,and then performs correlation calculation with the newly input frame.The obtained confidence map is the predicted tracking result,obviously,the point with the highest score(Or block)is the most likely trace result.A kernel function is used to map the low-dimensional space to a high-dimensional space.Make this algorithm have the following characteristics:1.SIFT feature is a local feature of the image.It has good invariance to translation,rotation,scaling,brightness change,occlusion and noise,and also maintains a certain degree of stability for visual changes and affine transformation.2.Multiplicity,even a small number of objects can produce a large number of SIFT feature vectors.3.The uniqueness is good and the information is abundant.It is suitable for fast and accurate matching in the massive characteristic database.4.The speed is relatively fast and can be optimized to meet real-time requirements.5.It is highly scalable and can be easily combined with other forms of feature vectors.In order to verify the effectiveness of the proposed algorithm,10 sets of complex video sequences were selected for testing and compared with the other 7 excellent tracking methods.The results show that the method proposed in this paper is 6.9% higher than the average distance accuracy of the best tracking method among the above-mentioned 4excellent tracking methods,and the complexity of the target changes in scale,lighting changes,partial occlusion,attitude changes,rotation,rapid movement,etc.The scene is more robust.