Siamese network visual tracking algorithm based on second-order attention

In order to improve the feature expression and discrimination capabilities of the Siamese network visual tracking algorithm and obtain better tracking performance, a lightweight Siamese network visual tracking algorithm based on second-order attention is proposed. First, use the lightweight VGG-Net as the backbone of the Siamese network to obtain the deep features of the target; then, use the residual second-order pooling network and the second-order spatial attention network proposed in this article in parallel at the end of the Siamese network to obtain features with Second-order attention features with channel correlation and second-order attention features with spatial correlation; finally, using residual second-order channel attention features and second-order spatial attention features, visual tracking is achieved through a dual-branch response strategy. The proposed algorithm is trained end-to-end using the GOT-10k data set, and verified on the data sets OTB100 and VOT2018. Experimental results show that the tracking performance of the proposed algorithm has been significantly improved. Compared with the benchmark algorithm SiamFC: on the data set OTB100, the accuracy and success rate have improved by 10.0%and respectively 9.6%; on the data set VOT2018, the expected average overlap rate (EAO ) was improved 7.7% ; and tracking speed reached 48 FPS.