Adaptive Hierarchical Siamese Network for Object Tracking

A new object tracking algorithm based on hierarchical Siamese network is proposed to solve the challenges of target deformation and background interference in tracking tasks. The introduction of hierarchical features makes up for the lack of features information caused by too small target. A similarity measure mechanism is firstly learned through matching training of extracted hierarchical features on annotated datasets. Online target tracking is achieved by searching candidate images at multiple scales around the target center of the previous frame. The adaptive examplar and the first frame exemplar are used to measure the similarity of the current frame to obtain an accurate score map, which allows the algorithm to adapt to the change of the target. Our algorithm achieves realtime target tracking without online fine-tune and achieves longterm effective tracking on multiple challenging video sequences.