引入再检测机制的孪生神经网络目标跟踪  

Siamese network tracking with redetection mechanism

作　　者：梁浩 刘克俭[3] 刘康 刘岩俊[1] 陈小林[1] LIANG Hao;LIU Ke-jian;LIU Kang;LIU Yan-jun;CHEN Xiao-lin(Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, China;University of Chinese Academy of Sciences, Beijing 100049, China;Remote Sensing Center, People′s Public Security University of China, Beijing 100038, China)

机构地区：[1]中国科学院长春光学精密机械与物理研究所,吉林长春130033 [2]中国科学院大学,北京100049 [3]中国人民公安大学,北京100038

出　　处：《光学精密工程》2019年第7期1621-1631,共11页Optics and Precision Engineering

基　　金：国家重点研发计划资助项目(No.2016YFC0803000);长春市科技发展计划资助项目(No.17DY008)

摘　　要：针对全卷积孪生神经网络SiamFC在目标快速运动、相似干扰较多等复杂场景下跟踪能力不足的问题,本文引入SINT作为再检测网络对SiamFC进行了改进。本文算法在跟踪响应图出现较多波峰时,启用精确度更高的再检测网络对波峰位置进行重新判定。同时,本文采用了生成式模型构建模板来适应目标的各种变化,以及高置信度的模型更新策略来防止每帧更新可能对模板带来的污染。在OTB2013上对算法性能进行了测试,并选取了9个主流的目标跟踪算法进行对比,本文算法的跟踪精确度达到了88.8%,排名第一,成功率达到了63.2%,排名第二,相比SiamFC有很大地提升。对不同视频序列的分析结果表明,本文算法在目标快速运动、严重遮挡、背景杂波、光照变化和长期跟踪等场景下具有较强的准确性和鲁棒性。To solve the insufficient tracking capability problem for a fully convolutional Siamese network (SiamFC) in complex scenarios such as those involving fast motion and large similar interference, SINT was introduced as a redetection network to improve the SiamFC. When multiple peaks appeared in the tracking response map, the proposed algorithm enabled the redetection network to redetermine the target position with higher accuracy. At the same time, a generative model was adopted to construct a template to adapt to various appearance changes of the target, and a high-confidence model update strategy was used to avoid the model corruption problem. Our algorithm is tested on OTB2013, and nine state-of-the-art algorithms are selected for comparison. The tracking accuracy of our algorithm reaches 88.8%, the best among all the algorithms selectes for comparison, and the success rate reaches 63.2%, which is the second best. Both these properties offer considerable improvement over the SiamFC results. Analysis of several representative video sequences demonstrate that our algorithm has high accuracy and strong robustness in cases involving fast motion, severe occlusion, background clutter, illumination changes, and long-term tracking.

关 键 词：目标跟踪 孪生神经网络 再检测 生成式模型 高置信度更新 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]