Siamese network tracker with attention mechanism

Abstract

The object tracking algorithm based on the deep Siamese network has a low tracking success rate and low robustness under the conditions of target illumination changes, occlusion and deformation. Therefore, this paper proposes a deep Siamese network tracking algorithm with attention mechanism based on SiamRPN++. First, an attention mechanism is added to each layer of the feature extraction network ResNet50 to calculate the importance of each channel, so that the model can obtain more useful information. Second, since the shallow features focus on the details of the target, the deep features focus on the semantic information of the target. Therefore, a feature fusion method based on the attention mechanism is proposed to fuse the deep and shallow features to enhance the expressive ability of the features. On the OTB100 and LaSOT datasets, the success rate of our tracker is 70.1%, 51.6%. Compared with SiamRPN++, it has increased by 1.1% and 2.4%.