Target tracking based on adaptive scale transform and feature fusion

Abstract

In order to achieve real-time stable tracking of moving targets and improve the accuracy and success rate of the tracking system, a kernel correlation filtering-based target tracking method with scale adaptation and feature fusion is proposed to address the situation that the traditional kernel correlation filtering algorithm does not track well when the target is obscured or motion blurred. Firstly, in the feature extraction process, color features are added after the original directional gradient histogram features to improve the recognition capability of target features, that is HOG features are fused with CN features, then a scale pyramid is constructed to perform scale estimation to achieve scale adaptation of the target, and finally the model is updated through a multi-peak detection mechanism. Through testing on the OTB2015 dataset, the accuracy and success rate of the algorithm has been further improved, and the algorithm is able to accurately identify targets and track them effectively.

Keywords

• target tracking technology|correlation filtering|feature fusion|scaling|multi peak detection