Siamese Visual Tracking with Correlation and Awareness

Target tracking has a wide range of applications in the fields of intelligent surveillance, human-computer interaction, and intelligent driving. In order to make the target tracking technique play a better role in these fields,an Siamese network target tracking algorithm (CDPSiamCAR) is proposed to address the problems that the traditional tracker generates background interference, which disturbs the localization of the target. The CDPCorr method is presented, which adds a dual-pixel cross correlation method to the traditional depth-wise cross correlation (DW-Corr), narrowing the sensory field to allow the tracker to focus more on the target itself, and also preserving the semantic information between the different channels. The Triplet Enhancement Attention Module is used prior to cross correlation to capture cross dimensional interactions and enhance the information matched by the cross correlation operation. The Regression Box Offset Module is introduced to be aware of the boundary information of the target and offset the original regression box to produce a more accurate regression box. The experimental results show that CDPSiamCAR produces improved results compared to the baseline model on both the OTB100 and UAV123 datasets, and excels in the challenges of background interference and scale transformation.