结合特征置信度的背景感知相关滤波跟踪算法  

Background-aware correlation filtering tracking algorithm combined with feature confidence

作　　者：胡国清[1] 陈辽林 刘谦波 戈明亮[1] JAHANGIR Alam SM HU Guoqing;CHEN Liaolin;LIU Qianbo;GE Mingliang;JAHANGIR Alam SM(School of Mechanical and Automotive Engineering,South China University of Technology,Guangzhou 510000,China)

机构地区：[1]华南理工大学机械与汽车工程学院,广东广州510000

出　　处：《现代电子技术》2021年第17期72-79,共8页Modern Electronics Technique

基　　金：广东省自然科学基金项目(2016A030313520)。

摘　　要：基于相关滤波的跟踪算法表现突出,然而在复杂跟踪环境下跟踪器很容易发生预测错误而导致跟踪失败。针对这一问题,提出一种结合特征置信度的背景感知跟踪算法。首先,采用较大的样本采集区域捕捉背景信息对滤波器训练,利用二次项替换及交替方向乘子法对模型进行快速求解。在颜色特征、方向梯度直方图特征以及灰度特征等多特征融合的基础上,根据目标响应大小对每个特征通道动态计算特征置信度,通过特征置信度加权来训练新的滤波器。为提高滤波器鲁棒性,结合APCE和特征置信度对滤波器进行高置信度更新。所设计的算法在OTB100数据集上与多个优秀相关滤波算法进行对比,实验结果显示,该算法在多种复杂场景下均具有较好的准确性和实时性。In recent years,the tracking algorithms based on correlation filtering have had an outstanding performance.However,the trackers are prone to occurrence of prediction errors in complex tracking environments,which results in tracking failures. In view of this,a background-aware tracking algorithm combined with feature confidence is proposed. A larger sample collection area is used to capture background information to train the filter. The quadratic term substitution and alternating direction method of multipliers(ADMM)are used to quickly solve the model. On the basis of the multi-feature fusion of color,histogram of oriented gradient(HOG) and gray,the feature confidence is dynamically calculated for each feature channel according to the degree of target response. The new filter is trained by weighting the feature confidence. In combination with APCE(average peak-to correlation energy)and feature confidence,the filter is subjected to updating for high confidence to improve its robustness. The designed algorithm is compared with a number of excellent related filtering algorithms on the data set OTB100.The experimental results show that the algorithm has good accuracy and real-time performance in a variety of complex scenarios.

关 键 词：滤波跟踪算法 背景感知 特征置信度 APCE 背景信息捕捉 特征融合 滤波器训练 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]