Abstract:

Kernelized correlation filter is one of the important algorithms in the field of visual object tracking. The object search region of this algorithm is limited. The rapid movement and abrupt direction change of the object may lead to tracking failure. To overcome this problem, “we introduce attention mechanism to improve the object tracking algorithm on the design of kernelized correlation filter. Firstly, the motion area is extracted through the bio-inspired retina model. Subsequently, the object candidate box is determined based on the average optical flow of the previous frame bouning-box. Finally, the object’s bounding-box is determined by kernelized correlation filter algorithm on the candidate box. Experimental results on the Anti-UAV2020 dataset show that when the PyrLK algorithm is used to calculate the optical flow, the tracking accuracy and success rate are improved by 1.4% and 1.3%, respectively, compared with the baseline method of kernelized correlation filter.When the Flownet algorithm is used to calculate optical flow, the tracking accuracy and success rate are improved by 2.2% and 1.3%, respectively, compared with the baseline method of kernelized correlation filter.

Key words:

Object tracking ,  , Correlation filter ,  , Optical flow algorithm ,  , Retina mode