Adaptive Feature Fusion Object Tracking with Kernelized Correlation Filters

Abstract

Aiming at the problem of occlusion in object tracking, adaptive feature fusion object tracking with kernelized correlation filters is proposed to advance the robustness of the correlation filter for object tracking. Firstly, feature fusion based on tracking confidence is proposed to improve the robustness of object tracking. Then combined with the detection method of tracking confidence to solve the object occlusion and tracking failure. Finally, template updating strategy based on tracking confidence is proposed to solve the drift problem. The experiment carries on the OTB100 dataset, and compares with other tracking algorithms. The results show that the tracking precision is 4.5% higher than that of the optimal algorithm, and the tracking success rate is improved by 6.1%. Moreover, the object can be well tracked under the occlusion.