| Object tracking with siamese region proposal network refers to the tracking method of deep learning.It can show better tracking results within the application scenarios of target tracing.But the method can’t extract the network through modern characteristics with both limited extracting ability and limitations of region proposal network.All these lead to a decrease in performance in such complex tracking situations as rotating targets and similar scenario.To address the issues mentioned above,the article has proposed Multi-stage Rotating Siamese Network Tracking.First of all,the introduction of modern polystage features aims to extract the network.The increase in depth of backbone network can enhance the ability of extracting features of backbone network.Meanwhile,backbone network can be divided into five stages with feature map processed in the last three stages,thus avoiding the loss of detail features to some extent.And then the introduction of Multi-stage Rotating Region Proposal Network can operate on the rotating regions of feature map at different levels and get a series of region proposal feature map.In doing so,it can allow trackers to inhibit the impacts of background on targets and improve target tracking effects in rotating scenarios and others.Last,sub-network of dual-lateral collaborative forecasting can forecast the location and rotating angel of targets so as to improve the coverage rate and tracking accuracy of final results.Compared to object tracking with siamese region proposal network,the algorithm presented in the article raises the accuracy rate by 6.4% and success rate by 8.5% in terms of OTB2015 data set.As for VOT2016 data set,the accuracy rate increases by 4.1%,and average coverage rate by 1.4%,but the former increases by 9.9%,and average coverage rate by 3.4% as for VOT2018 data set.At the same time,the proposed algorithm demonstrates strong robustness and adaptability to multiple complex conditions such as rotation,deformation and illumination.The tracking speed is up to 86.8 fps,meeting real time requirements.This paper has 30 figures,5 tables,and 70 references. |
