Visual Tracking Method Based on Siamese Network with Multi-Feature Fusion

The traditional deep learning tracking method SiamFC faces performance degradation while solving issues, for instance, similar background, occlusion, target deformation, and illumination variation. This paper proposes an improved SiamFC with multi-feature fusion strategy. The proposed method first extracts the histogram of gradient and color name of the template image and search area by correlation filter. Then, the method fuses them and weights the SiamFC response map to obtain a more accurate object response position. Comparison experiments on VOT and OTB datasets prove that the improved method is more accurate and robust than the excellent tracking methods to deal with problems such as target cover, out of sight, scale variation and motion blur.