A Siamfc Target-Tracking Algorithm Based on an Improved Spatiotemporal Attention Mechanism

Abstract

The motion target tracking algorithm has developed rapidly. In this paper, the SiamFC (Fully Convolutional Siamese Networks) algorithm mainly relies on the first frame of the video as a template, and lacks an effective update mechanism. Based on the SiamFC algorithm, This paper  introduced an improved spatiotemporal attention mechanism, and the model pays more attention to key historical frames and target regions in the video sequence by introducing an improved spatiotemporal attention mechanism in the backbone network. Moreover, the pixels in the response map are divided between background and foreground by a pixel-by-pixel classification regression method. By combining the centrality branch to limit the generation of lower quality prediction box, increase the accuracy of target prediction and reduce the complexity of prediction, the algorithm improves the accuracy and success rate, effectively realizing the target tracking in complex scenarios, while maintaining the accuracy and stability of tracking.