Long-Term Visual Tracking of an Unmanned Surface Vehicle in Complex Environment Based on Adaptive Multi-Feature Fusion with Correlation Filtering

Abstract

In this paper, a long-term visual tracking scheme of an unmanned surface vehicle (USV) based on adaptive multi-feature fusion is proposed for visual tracking in complex environment. The salient features of the proposed approach are: (1) A scale filter is used to solve the problem of performance degradation caused by scale change of the USV in visual tracking and the dimension reduction strategy is applied to the scale filter to accelerate scale detection in the tracking process. (2) To alleviate problem such as illumination variation, deformation and background color similarity in complex environment, the Histogram of Gradient (HOG) and Color Names (CN) features are successfully combined to improve reliability and accuracy of the visual tracking scheme. (3) The anti-interference ability of visual tracking is enhanced by training Support Vector Machine (SVM) re-detection module online in order to eliminate external factors such as occlusion and moving out of sight during long-term visual tracking thereby strengthening robustness of the visual tracking scheme. Experimental results and comparative studies demonstrate that the proposed scheme is superior in tracking.