Research On Object Tracking Algorithm Based On Regularization

Object tracking has made outstanding achievements in military,video surveillance,behavior analysis and other fields.This model uses the information of the first frame of the learned sample to identify and locate the target in the next consecutive frames.The correlation filter algorithm uses a ridge regression model composed of input samples and ideal Gaussian weighted responses to fit the best target template.Due to the periodic repetition of cyclic shift samples,the performance of model is degraded.Many existing methods can be used to improve the accuracy and speed of the model,such as multi-channel features,kernelized filters,scale-adaptive,and statistical learning methods.Although these methods have achieved some success,the video attribute challenges faced by target tracking cannot be solved.Therefore,this paper focuses on improving the performance and speed of the target tracking model and effectively combining the optimization methods in machine learning to further improve the robustness and reliability of the target template.The specific research content is as follows:(1)The method based on L2 regularization has achieved the best fit of the filter coefficients between the sample and the Gaussian weighted response.To improve the convergence speed,this dissertation introduces the L1 constraint of the filter coefficients.L1 regularization helps to generate a sparse weight matrix,which can be used for feature selection,indicating that only a few features contribute to this model.Most of the features are not of much significance,we only pay attention to the non-zero coefficients feature.The L1-STRCF model proposed in this paper is tested on the OTB-2015 and La SOT datasets.The experimental results show that the tracking speed of the model is 13% higher than the original,and it is more robust when it encounters video challenges such as occlusion,deformation,and fast motion.(2)Aiming at the problem that the target background of the correlation filter is not modeled over time,which leads to sub-optimal performance,this paper introduces the time regularization term of the background cropping operator based on the scale adaptive method.Using the sample information extracted from the real background to expand the training samples,the negative samples generated by real shifts expand the search area,avoiding boundary effects,making the learned coefficients more reliable and the classification more accurate.The time regularization can remember the position information of the previous target,and relocate the target during occlusion and motion blur to achieve the best match.This paper uses the ADMM algorithm to achieve the derivation of the model,optimizes the model into several related sub-problems,and solves it in alternative directions iteratively and guarantees that the optimal solution after 2 iterations,reducing the computational complexity.Finally,experiments on the OTB-2015 data sets prove that the improved model proposed in this paper is higher in accuracy and speed than several currently popular algorithms,ranking first in comparison experiments.(3)The tracking model has been using the ridge regression method to fit the best coefficients in order to improve performance.The paper called ASRCF published on CVPR2019 began to use weight regularization to improve the model,and used the prior weight wr to achieve weight adaptation,ASRCF has a specific weight coefficient for different target sequence,this idea is very different from previous work.Drawing on the success of L1 regularization in the field of face recognition,this article introduces the L1 norm term based on spatial weights.Compared with filter-based optimization model,one is the template itself,and the other is the optimization of other parameters.After adding sparse constraints,the obtained target scale feature map has no edge burrs,the features displayed are more careful,the principal component representation is more complete,and the model performance is more excellent.The research work in this article is simulated on the OTB-2015 dataset.The scale feature map is more representative than the original.It ranks first in speed and accuracy among several current trackers,and obtains the optimal solution after 2 iterations.This paper researches and improves the target tracking model based on correlation filter,and introduces the optimization terms of time regularization,spatial weight constraint and L1 regularization for the model.Then,we use ADMM algorithm to optimize the model,and combine the new constraint idea to make the model real-time and reliable,and more business prospects.