《计算机应用研究》|Application Research of Computers

结合学习率调整的自适应特征融合相关滤波跟踪算法

Adaptive feature fusion correlation filter tracking algorithm combined with learning rate adjustment

免费全文下载 (已被下载 次)  
获取PDF全文
作者 成悦,李建增
机构 陆军工程大学石家庄校区 无人机工程系,石家庄 050003
统计 摘要被查看 次,已被下载
文章编号 1001-3695(2019)07-061-2210-04
DOI 10.19734/j.issn.1001-3695.2018.01.0121
摘要 针对单一特征存在的缺陷和目标快速变化时易跟丢的问题,提出了一种结合学习率调整的自适应特征融合相关滤波跟踪算法。采用互补的梯度特征和颜色特征进行特征融合,通过计算滤波响应的大小来决定下一帧在融合特征中各自所占的权重,凸显优势特征,使目标与背景更具区分度;提取目标后需要更新滤波器,为了避免滤波器跟不上目标变化的情况发生,引入学习率调整机制,使滤波器更新速度能够随目标外观变化进行在线调整。相较同类特征融合算法,提出的算法准确高效,且对于快速形变目标的鲁棒性更强。实验证明,该算法在精度和成功率上都比现有相关滤波算法更优,具有一定的应用价值。
关键词 目标跟踪; 相关滤波; 特征融合; 自适应加权; 学习率
基金项目 国家自然科学基金资助项目(51307183)
本文URL http://www.arocmag.com/article/01-2019-07-061.html
英文标题 Adaptive feature fusion correlation filter tracking algorithm combined with learning rate adjustment
作者英文名 Cheng Yue, Li Jianzeng
机构英文名 Dept. of UAV Engineering,Army Engineering University,Shijiazhuang 050003,China
英文摘要 In view of the defects of single feature and the problem of easy to miss when the target is fast changing, this paper proposed an adaptive feature fusion correlation filter tracking algorithm based on learning rate adjustment. The algorithm used complementary gradient features and color features to fuse the feature, and decided the weights of the features in the fusion feature in next frame by calculating the size of the filter response of each feature, so as to highlight the dominant features and make the target and background more discriminative. After extracting the target, it needed to update the filter. In order to avoid the situation that filter couldn't keep up with the change of the target, this paper introduced the learning rate adjustment mechanism, so that the update speed of the filter could be adjusted online with the appearance of the target. Therefore, compared with the similar feature fusion algorithm, the proposed algorithm is more accurate and efficient, and the robustness of the fast deformation target is stronger. The experiment shows that this algorithm is better than the existing correlation filter algorithms in accuracy and success rate, and has a certain application price.
英文关键词 target tracking; correlation filtering; feature fusion; adaptive weighting; learning rate
参考文献 查看稿件参考文献
 
收稿日期 2018/1/31
修回日期 2018/3/20
页码 2210-2213
中图分类号 TP391.41
文献标志码 A