An Improved Staple Algorithm With Attention Mechanism

Since the traditional Staple target tracking algorithm has poor robustness and precision in complex and changeable scenes, this paper proposes an improved Staple algorithm with the attention mechanism. Specifically, to cope with the variability of different scenes, the weights of the template feature response map and the histogram feature response map are adaptively allocated by introducing the attention mechanism, which is used to obtain the fused response map that combines the advantages of two feature map. In addition, the VGG-Net depth feature for target position correction is introduced, which is linearly superimposed with the fused response map to achieve more accurate target positioning. On the two datasets of OTB50 and OTB100, the improved Staple algorithm is qualitatively and quantitatively compared with other methods, and the experimental results show that this algorithm achieves more accurate target positioning and tracking.