In the application environment of railway traffic scenes,object tracking is easily affected by background clutter,intense object motion,scale transformation and so on.It is easy to lose the object due to inadequate object tracking accuracy.Therefore,an object tracking algorithm based on twin networks is presented for complex railway scenes.Firstly,in the feature extraction phase,CIResNet-22,which is an improved ResNet network,is used as the feature extraction backbone network,replacing the naive correlation with pixel-wise correlation.Secondly,a normalization-based attention module is added,which enables the network to focus on reliable features while weakening unreliable ones.Finally,the results reflected by the average peak correlation energy are used to determine whether the tracking results are reliable,and the best template predicted by an improved UpdateNet subnetwork is used as a reference template.The results show that good tracking performance can be obtained on standard datasets such as VOT2018,VOT2016 and OTB100.At the same time,the tracking sequence is tested in the self-made video sequence with good results.
维普
万方数据