Research On Object Tracking Based On Siamese Neural Network

Object tracking is one of the most popular and challenging research field in computer vision,and has a wide range of applications in areas such as autonomous driving and surveillance.In recent years,deep learning based trackers have drawn great attention in object tracking because of their deep architecture,discriminative deep features.SiamRPN,inspired from the object detection,combines the Siamese Neural Network and Region Proposal Network and have achieved state-of-the-art performance with real-time speed.However,SiamRPN uses Alex Net as the backbone network which restrict the feature extractor of SiamRPN since Alex Net is not deep enough,the deep features extracted by SiamRPN are not discriminative enough to effectively use semantic information.SiamRPNCA is proposed to address this problem which combines the attention mechanism and SiamRPN.It is designed to learn effective feature representation.Experimental results on three benchmarks,OTB50,OTB100 and VOT2016,show that SiamRPN-CA has performed better than SiamRPN.In addition,SiamRPN-CA is extended for long-term tracking,and experiments are performed on long-term tracking dataset VOT2018 LT and show that this approach has significantly improved the long-term tracking ability.The main research contents include:1.Research and analyze about SiamRPN.Based on the result,a short-term object tracking approach named SiamRPN-CA which combines SiamRPN with a channel attention block which can improve the ability of feature extractor is proposed.Experiments performed on OTB50,OTB100 and VOT2016,indicates the significantly improvement on SiamRPN-CA.2.SiamRPN-CA is extended for long-term tracking.Experiments are performed on VOT2018 LT and the results verifies the improvement.