Research On Object Tracking Algorithm Based On Siamese Convolutional Neural Network

In recent years,object tracking has become one of the research hotspots in the field of computer vision.Since the object tracking algorithm based on the siamese convolutional neural network has been proposed,it has attracted the attention of a large number of researchers by taking advantage of both speed and accuracy.At the same time,this type of algorithm still faces many challenges.When the actual environment is complex,tracking drift or failure is prone to occur.In order to further improve the accuracy of the algorithm under different challenging factors,this thesis firstly improves the algorithm template setting and search range selection strategy to improve the accuracy of the algorithm in fast-moving,similar background scenes.However,due to the insufficient generalization ability of the single feature extraction network structure,the accuracy of the algorithm in other parts of the challenge is limited.Therefore,an online selection multi-branch model is proposed by this thesis subsequently.The algorithm has good stability in complex scenes such as motion blur and lighting changes.The research results and main contributions of this thesis are as follows:1.A siamese convolution tracking algorithm based on optical flow prediction is proposed.The current mainstream siamese convolution algorithm uses the object of the first frame as a template.When the appearance of the object changes in subsequent frames,drift is easy to occur.Moreover,the search range of the current frame is selected as the center of the object of the previous frame,which is not robust to the instantaneous long displacement scene of the object.To this end,the following improvements have been made to the selection of templates and search ranges: First,a template update mechanism is proposed to learn the appearance transformation parameter matrix online so that the result of the cyclic convolution with the template feature map is similar to the object feature map of the previous frame.The process is modeled using regularized linear regression,and the process shifts to the frequency domain solution.Secondly,an optical flow motion information prediction module is proposed,which uses the optical flow network to estimate the relative motion trend of the object between adjacent frames.The search range of the frame is selected according to the prediction result,so that the object is close to the search center,and the influence of the cosine window on the imbalance of the response map penalty is reduced.In order to prove the performance of the proposed template update mechanism and optical flow motion information prediction module,this algorithm is compared with other algorithms on public data sets.Experimental results show that,compared with the basic algorithm Siam Fc,the accuracy of this algorithm is improved by 4.1% on fast-moving scenes and 10.0% on similar background scenes.The algorithm has good robustness in situations such as fast motion,background interference,and deformation of non-rigid object.2.A siamese convolution tracking algorithm for multi-branch online selection is proposed.At present,the mainstream siamese convolution algorithm uses a single matching function,and does not update it online,resulting in poor generalization ability of the algorithm,which cannot effectively characterize the object features of some scenes.In response to this problem,a multi-branch structure is proposed by this thesis,the branch is divided into three parts:the first part is a common branch,its structure is consistent with Siam Fc;the second part is a diversified branch group,including 10 vertical Siam Fc branches,each branch is pre-trained with ILSVRC-2015,and then fine-tuned with the data after the training set clustering,so that each branch has the characterization ability of a specific scene,where the fine-tuning uses a tripletwise training strategy;the third part is the Res Net branch,This branch uses classification tasks for network pre-training.The feature extraction network is adjusted based on the18-layer Res Net structure.At the same time,a feature fusion mechanism is added to avoid the problem of reduced tracking accuracy due to the deepening of the network.When tracking online,select the branch with the highest discrimination in the response map to track,and select the branch again at 8-frame intervals.In order to prove the role of each branch in the multi-branch network structure,the ablation experiment was first carried out,and the results showed that each branch improved the performance of the proposed algorithm.Finally,compared with other mainstream algorithms on the OTB platform,experiments results show that the algorithm can effectively deal with complex scenes such as motion blur and lighting changes.