SiamDSF: An Improved Siamese Network Based on Deep and Shallow Feature Fusion for Target Tracking

Siamese trackers are gradually replaced by Transformer trackers with huge models, complex calculations, and redundant parameters. However, in industrial applications, Siamese trackers have a simple structure and lightweight design. It is more suitable for practical deployment. We believe that the Siamese tracker still has huge room for improvement. The SiamRPN++ tracker has received widespread attention due to its good balance between speed and performance. However, in complex backgrounds, SiamRPN++ often loses its target. It lacks the ability to recognize interference from similar objects when using deep semantic features for tracking. For this, we have designed a deep and shallow feature fusion module(DSFF). The fusion module cleverly uses inverted residual mobile block (iRMB) to capture global feature semantic information, and then relies on the feature fusion module (FFM) to fuse and filter shallow and deep features. It improves the generalization ability of the model and avoids the impact of similar targets in complex scenes that may lead to incorrect tracking. In addition, we widely use depthwise separable convolution to reduce the number of parameters and improve tracking efficiency. This tracker has achieved good results on different tracking benchmarks, especially on the UAV123 dataset. It can be applied to drone tracking and enemy search.