Object Tracking in UAV Videos by Multi-Feature Correlation Filters with An Adaptive Update Strategy

Abstract

Object tracking in UAV videos has attracted great attention in various fields, such as road monitoring, equipment inspection, and wildlife census. However, compared with traditional object tracking, it also faces severe challenges due to scale variation, low resolution, and partial occlusion, which have serious effects on the performance of the tracker. To settle those difficulties, the multi-feature correlation filters with an adaptive update strategy is proposed. First, we fuse histogram of oriented gradient (HOG) features, saliency (SA) features and color names (CN) features to enhance the discrimination ability of the tracker. Then, we use the denoising mechanism to mitigate the impact of similar background. Finally, we introduce the adaptive update strategy to prevent model degradation. Through experiments on multiple datasets, our algorithm can accurately track targets in UAV videos.