Research On Target Tracking Algorithm Based On Machine Learning

Target tracking is a hot research direction in the field of computer vision,which has important application value in intelligent security,autopilot and so on.In order to deal with the tracking problems in complex scenes such as occlusion and similar background,this paper studies the target tracking algorithm based on machine learning and proposes two algorithms to improve the accuracy and robustness of target tracking.The research contents of this paper are as follows:1.Aiming at the complex tracking scenes such as serious occlusion and fast deformation,a correlation filtering target tracking algorithm based on multi-feature fusion is proposed.Through the multi-feature fusion strategy,the extracted directional gradient histogram(f HOG)features and color(CN)features are fused.The tracking accuracy and robustness of scale changing targets are improved through the scale updating mechanism of position filter and scale filter.At the same time,a high confidence model update strategy is introduced to stop updating the filter model when the target is occluded,which reduces the probability of model drift when the target is occluded.In the open data set OTB-100 and VOT2018,the proposed algorithm is compared with other advanced algorithms,and the proposed algorithm has better performance.2.Aiming at the complex situations such as background similarity and rotation,a twin network target tracking algorithm based on self-clipping residual network is proposed.The sample frame and the current frame are inputted into the Res Net-34 self-clipping residual network constructed by the self-clipping residual unit(CIR),respectively,and the deep semantic features and shallow visual features are fused to improve the representation ability of the tracked target features.A rectangular boundary box is constructed to generate a network including the classification branch of the foreground and background and the regression branch of the rectangular box,which distinguishes whether it is a tracked target by the classification branch,and determines the center coordinate and edge length of the rectangular box by the regression branch of the rectangular box.Experiments are carried out on the open data sets OTB-100 and VOT2018,and compared with other advanced algorithms,the siamese network target tracking algorithm proposed in this paper has better performance.