光学学报, 2018, 38 (5): 0515003, 网络出版: 2018-07-10
基于分类-验证模型的视觉跟踪算法研究
Visual Tracking Algorithm Based on Classification-Validation Model
机器视觉 目标跟踪 分类-验证网络模型 类内相似 类间差异 machine vision target tracking classification-validation model intraclass similarity interclass differences
摘要
针对相似度目标跟踪算法主要考虑目标的类内相似,而忽略不同目标的类间差异的问题,提出基于分类-验证模型的视觉跟踪算法。该算法通过增加目标的属性(类别)信息,利用相似度信息与类别信息构建损失函数,在高维空间学习目标的类内相似和类间差异;将目标模板与候选目标输入网络模型,分别通过分类与验证模块实现网络参数更新;利用训练网络提取目标模板与候选目标的深度嵌入特征,实现目标跟踪。在OTB50和UAV123数据库上进行实验,结果表明,该算法可以大幅提高跟踪效果,对相似目标具有较强的稳健性。
Abstract
In order to solve the problems that the similarity target tracking algorithm mainly considers the intraclass similarity of targets and ignores the interclass differences of different targets. A visual tracking algorithm based on classification-validation model is proposed, which adds attribute information to the similarity algorithm. The proposed algorithm constructs the loss function with similarity and class information, and learns intraclass similarity and interclass differences in high dimensional space. The classification and verification module is adopted to update network parameters when the target template and candidate target input into the network model. With the trained network, the deep embedding feature of target and candidate target is extracted, thus, the target tracking is achieved. Experiments are carried out on the OTB50 and UAV123 databases. Results show that the proposed algorithm can improve the tracking effect with increased target information, and has strong robustness to the similar targets.
吴敏, 查宇飞, 张园强, 库涛, 李运强, 张胜杰. 基于分类-验证模型的视觉跟踪算法研究[J]. 光学学报, 2018, 38(5): 0515003. Min Wu, Yufei Zha, Yuanqiang Zhang, Tao Ku, Yunqiang Li, Shengjie Zhang. Visual Tracking Algorithm Based on Classification-Validation Model[J]. Acta Optica Sinica, 2018, 38(5): 0515003.