Research On Real-time And Robust Object Tracking Based On Correlation Filter And Siamese Network

As one of the most basic research fields of computer vision,visual object tracking is widely applied to different places,such as video surveillance,intelligent transportation,humancomputer interaction,medical diagnosis,assistance driving,and visual navigation.It's only given the center position and the size of any-interest object in the first frame,the task of object tracking is to continuously estimate the position and size of the object in subsequent video or image sequences in order to provide data information for other vision tasks.In recent years,researchers from all over the world have already been doing lots of studies on visual object tracking and getting relatively better tracking accuracy in simple scenario(such as rigid objects in static scenes).However,the object in actual complex scenes will be interfered by different factors such as illumination change,fast motion,occlusion,background clutter,it makes more difficult to model the appearance of the object.Therefore,how to achieve object tracking tasks accurate,efficient and robust is still highly challenging.Based on the analysis of the existing excellent tracking algorithms,this paper mainly aims to use the basic theory of related filtering and deep learning methods as the algorithm framework,and focus on establishing a reliable appearance model for more effectively in-depth studies of suppression model in degradation under complex and changeable scenario.The specific work and innovations of this paper are as follows:1)A tracking method that satisfies faster tracking speed and higher robustness in object tracking under complex scenes is proposed.A cost function is constructed by using all previous frames and the current frames in the whole samples,and the weighted average quadratic minimization is used to train the classifier.Based on the classifier,increasing the training samples can lead to improve the robustness of the classifier.In order to improve the time efficiency of the tracker,there are three different strategies.The first one shows Gaussian radial basis function kernel method and the Fast Fourier Transform technology will be applied based on the correlation filtering theory by using Gaussian radial basis function kernel method to reduce time cost;the second one is to carry the key information by using the object feature dimensionality reduction method when the object feature dimensionality is reduced.In this way,it proved that the calculation time cost can be greatly reduced and the effectiveness of the classifier can be maintained;The third one is to effectively perform the interpolation of the triangular polynomial in the detection response score,and use the sub-grid interpolation method to calculate the pixel-intensive response score to avoid the influence of the pixel-level calculation on the calculation amount.The above-mentioned method of improving calculation efficiency is also used in the object scale estimation to adapt to the change of the object scale.In addition,the classification ability of the classifier is strengthened by expanding the search area and adding more positive and negative samples;In terms of model updating,the numerator and denominator based on the tracker are synchronously interpolated and updated to update the classifier by storing the currently model after learning.2)A tracking method based on weighted kernel correlation filtering is proposed.Kernelrelated filters and color histogram models construct a robust appearance model through parallel learning of spatially regularization: firstly,in order to reduce unnecessary boundary effects in the sample,spatial regularization components are introduced to learn kernel-related filters,the regularization weights of different samples can reduce the influence of boundary effect by penalizing correlation filter and keeping the robustness of the tracker;secondly,considering the color information in the picture,the spatial reliability map constraint training samples are constructed by using the color histogram model.Moreover;two complementary features including hog and Hoi descriptors are integrated to construct a robust and reasonable multichannel object feature description and improve the discrimination ability of the classifier.Finally,the feedback mechanism of tracking results is used to determine the way to update the online model.3)The fixed weight of the filter cannot adapt to all different objects and changing scenes during the tracker process.The degradation of the filter can be restrained by strengthening the object feature description and the sequential information of continuous frames.In order to solve these problems,a correlation filter visual tracking method based on depth feature and spacetime constraint is proposed.The hand-crafted and shallow features of different convolutional neural networks are fused to describe the object features.In the framework of correlation filtering,the temporal information of continuous frames and the spatial information of current detected frames are fully considered when constructing the objective function of the classifier.The adaptive spatial weight constraint and temporal consistency constraint are introduced,and the ADMM technology is used to optimize the classifier.The advantage of the method is that there is no need to update the model of the classifier except the object feature.The combination of multi-feature fusion from different convolution neural network can greatly improve the classifier's ability to locate the object.At the same time,the above method can solve the problems of object scales only manual features.4)A Siamese network object tracking method based on online dynamic update is proposed.In the process of object tracking based on deep learning framework,the model structure of network is consistent with the structure in Siamese-FC tracker.In order to get the model parameters,the information from positive-negative samples is fully needed and trained by using three component loss instead of logistic regression loss;Considering the change of object shape,all the object templates in the previous frame to learn the transformation parameters based on the similarity of object appearance are used.There are two parameters capturing from the transformation parameters after learning based on background suppression through adding the Gaussian weight distribution in the detected frame.Using the above two parameters,the detection samples in the search area compared with the objects in the previous frame through the similar transformation of the object appearance to ensure object tracking.This method improves the performance of the tracker without sacrificing time efficiency performance.