| Target tracking has always been one of the important research directions in computer vision,because it can not only provide the moving state of the target,but also provide important data for high-level visual analysis such as scene understanding.At present,it has been widely used in human-computer interaction,intelligent video monitoring,public security and many other fields.The so-called target tracking refers to selecting the target of interest in an image or video sequence,finding the position of the target in the next consecutive frame and forming a motion track.Although with the development of deep learning,the target tracking algorithm has achieved many research results,but because of the influence of complex factors such as complex illumination changes,interference of similar objects in the background and camera jitter,it makes the target apparent model dynamic change,so it is still a very challenging task to study the target tracking algorithm which can effectively adapt to the dynamic changes of complex scenes The task.Based on the analysis and summary of the existing typical target tracking algorithms,this paper concludes that the existing target tracking algorithms are generally divided into two categories:(1)target tracking algorithm based on correlation filtering;(2)target tracking algorithm based on depth learning.The tracking algorithm based on correlation filtering has significant advantages in computational efficiency and competitive effect,but it has two main disadvantages(1)manual feature extraction can not capture the semantic information of the target;(2)lack of training data.In order to overcome the shortcomings of CF(correlation filter,CF)manual feature extraction performance,the depth convolution feature is introduced into CF,and the defect is successfully optimized.Although the CF target tracking based on the characteristics of the convolution neural network(CNN)has overcome the geometric and environmental changes to a certain extent,and improved the robustness.However,the computation of extracting CNN features from each frame and training/updating CF target tracking is still very large.Therefore,it is still very difficult to achieve real-time target trackingIn order to solve the problem of large amount of computation in extracting CNN features,this paper improves on the basis of full convolution symmetric network,and combines with correlation filter,adds fast Fourier transform(FFT)to reduce the computation.CF can distinguish a block image from the surrounding blocks by solving the ridge regression problem efficiently.Because of FFT and element by element operation,CF is more efficient than SGD algorithm.Compared with the embedded method,the discriminator is more suitable for specific scene target.Based on the previous research work,the process of applying CF to the pre trained CNN features is deduced,and the end-to-end combination of the two methods is realized.The key point is to use CF as a differentiable CNN layer,so that the error can be transmitted back to CNN features through CF.Firstly,in each video scene,the reference area block which can distinguish the appearance of the target is selected in advance to construct the training sample,and two incomplete symmetrical convolution networks are constructed.The convolution network makes the output characteristics of the target outside the reference region similar to those of the target inside the reference region as much as possible,so as to obtain a good representation of the target inside the reference region.In one of them,the correlation filtering module is added.Finally,the end-to-end deep learning and correlation filtering framework is built.First,the algorithm is compared with SiamFC.When the shallow network computing features are used,the algorithm in this paper is obviously superior to SiamFC.Then,the simulation experiments with different network depths are carried out with SiamFC and DSST.The experimental results show that for the deep enough symmetrical network structure,the end-to-end deep learning and correlation filtering framework is added The correlation filter layer does not significantly improve the tracking accuracy,but if there is enough training data,the ability of deep learning to extract strong representation cannot be underestimated.Figure[45]table[4]reference[84]... |
