欢迎访问《图学学报》 分享到:

图学学报

• 计算机视觉 • 上一篇    下一篇

特征融合自适应目标跟踪

  

  1. 南昌航空大学计算机视觉研究所,江西 南昌 330063
  • 出版日期:2018-10-31 发布日期:2018-11-16
  • 基金资助:
    国家自然科学基金项目(61663031,61661036);江西省重点研发计划项目(20161BBE50085,20171ACE50024);南昌市优势科技创新团队 建设项目;江西省教育厅科学技术项目(GJJ150738)

Feature Fusion Adaptive Target Tracking

  1. Institute of Computer Vision, Nanchang Hangkong University, Nanchang Jiangxi 330063, China
  • Online:2018-10-31 Published:2018-11-16

摘要: 经典视觉单目标跟踪方法通常以单特征描述被跟踪的目标。但在实际场景中,目 标因受外界因素如光照或自身变化如形变的影响而发生变化。为了更好地描述目标,首先引入 HOG 特征和 CN 特征,利用传统的特征提取方法,训练得到各自的相关滤波器;然后与各自特 征相关滤波得到各自的响应图;最后采用实际响应与期望响应的差值法求得各自响应图的权重, 将其与各响应图自适应融合得到目标的最终位置,并自适应更新各自的模型。实验选取公共数 据集 OTB2013 的 34 个彩色视频帧序列对不同算法进行定性和定量地分析和论证。相比效果最 好的 DSST 算法,平均中心误差减少了 7.8 像素,成功率提高了 1.2%,精度提高了 2.3%。实验 结果表明该算法具有较好的跟踪鲁棒性和准确性。

关键词: 目标跟踪, 相关滤波, 权重, 特征融合, 模型自适应更新

Abstract: The classical visual single-target tracking method usually describes the tracking target with a single feature. But in the actual scene, the target is changed by the external factor such as light or its own changes such as deformation. In order to better describe the target, the HOG feature and CN feature are introduced firstly, and the traditional feature extraction method is used to train the respective correlation filter, and then the response graph is obtained by correlating filter with the respective feature. Finally, the weights of each response graph are obtained by using the difference method between actual response and expected response, and the final position of the target is obtained by adaptively fusion with each response graph, and their models are updated adaptively. The experiment selected 34 color video frame sequences from the common data set OTB 2013 to analyze and prove the different algorithms qualitatively and quantitatively. Compared with the best DSST algorithm, the average center error is reduced by 7.8 pixels. The success rate is improved by 1.2% and the accuracy is improved by 2.3%. The experimental results show that the algorithm has good tracking robustness and accuracy.

Key words: target tracking, correlation filter, weight, feature fusion, model adaptive update