Object Tracking Algorithm Based On Few-shot Learning

With the rapid development of deep learning,in recent years,single object tracking based on deep learning has been greatly improved in terms of algorithm performance,and this part of the improvement is largely due to the good resolution ability of deep features that extracted by deep neural networks.Object tracking has the following problems in practical application.Firstly,the few-shot data is insufficient and the sample diversity is low,which leads to the limited number of features extracted from the network.Secondly,the object tracking accuracy is low under the interference of occlusion,deformation and fast movement.The main research directions and innovations are summarized as the following three points:(1)In order to solve the problem of insufficient few-shot data and low sample diversity,two data enlargement strategies are proposed to expand the sample and improve the sample diversity.(2)By establishing a nonlinear motion CS model and introducing multiple fading factors to correct the estimated object motion state,the problem of the tracking performance decline of volumetry Kalman filter after the mutation of object state is solved.This method not only effectively solves the problem of filter performance degradation when the object motion state changes,but also improves the tracking accuracy and adaptability of the model when the object motion state changes.The root mean square error of the center point of the experiment is reduced by 4.84 and 6.62 respectively in the Basketball and Biker sequences of OTB2015,and by 6.43 in the remote sensing images.(3)Aiming at the problem that Siamese networks can produce high response to objects with similar semantics in the search region,an object tracking algorithm based on location information for Siamese region proposal network is proposed.Combined with the innovation point(2),the object motion state is used to predict the object position information and the edge penalty is introduced to suppress the high-score interference items far from the object center in the search area,which effectively solves the interference problem that similar semantic objects will produce high response.The average success rate of the experiment increased by 6.7% in OTB2015 and 5.3% in remote sensing images.The algorithm proposed in this paper is compared with the representative Siamese network and related filtering object tracking algorithms on the OTB2015 data sets.The experimental results show that the algorithm in this paper is superior to the experimental comparison algorithm in terms of average performance and robustness.The feasibility of the algorithm is verified.