胡昭华, 李奇, 韩庆. 基于双检测器系统的长期目标跟踪算法[J]. 微电子学与计算机, 2021, 38(9): 31-37.
引用本文: 胡昭华, 李奇, 韩庆. 基于双检测器系统的长期目标跟踪算法[J]. 微电子学与计算机, 2021, 38(9): 31-37.
HU Zhaohua, LI Qi, HAN Qing. Dual-detector system for Long-term tracking[J]. Microelectronics & Computer, 2021, 38(9): 31-37.
Citation: HU Zhaohua, LI Qi, HAN Qing. Dual-detector system for Long-term tracking[J]. Microelectronics & Computer, 2021, 38(9): 31-37.

基于双检测器系统的长期目标跟踪算法

Dual-detector system for Long-term tracking

  • 摘要: 针对长期目标跟踪过程中极易出现的目标漂移导致跟踪失败的问题,在空间正则化相关滤波算法的基础上,提出一种基于双检测器系统的长期目标跟踪算法.为了适应不同的跟踪场景,增加模型鲁棒性,加入SVM检测器和孪生网络检测器,并选择较为高效的SVM检测器作为主检测器;只有当主检测器检测失败时,才启动精度更高的孪生网络检测器检测目标,通过两个检测器的协同工作可以对当前图像进行高质量的重新检测.同时,本算法中加入自适应空间正则化项,可以有效抑制边界效应的影响;并采用交替方向乘子法对目标函数进行优化,降低计算复杂度以节约运行时间.本文算法在OTB-2013和OTB-2015两个数据集上进行测试,结果表明:本文算法在多种特性上都有较强的鲁棒性,特别是在进行长视频序列的跟踪时,比基础算法在跟踪距离精度和成功率上分别提高了12.9%和6.2%.

     

    Abstract: Aiming to solve the problem of target drift that occurs often in the process of long-term tracking, a dual-detector system for long-term tracking is proposed based on the spatially regularized correlation filters algorithm. In order to adapt to different tracking scenarios and increase the robustness of the model, SVM detector and siamese network detector are added, and the more efficient SVM detector is selected as the main detector. Only when the main detector fails to detect the target, the more accurate siamese network detector is started to detect the target. Through the cooperative work of the two detectors, the current image can be re-detected with high quality. At the same time, the adaptive spatial regularization term is also added in this paper, which can effectively suppress the influence of the boundary effect. The objective function is optimized by the alternating direction multiplier method to reduce the computational complexity and save running time. The experiment on datasets OTB-2013 and OTB-2015 show that the proposed algorithm has strong robustness in many characteristics. Especially in long-term tracking, the tracking distance accuracy and success rate of the proposed algorithm are improved by 12.9% and 6.2% respectively compared with the basic algorithm.

     

/

返回文章
返回