基于上下文感知的自适应目标跟踪算法  

Target Tracking Algorithm Based on Adaptive Context Awareness

作　　者：万文明 路红[1] 秦彬鑫 邱春 WAN Wen-ming;LU Hong;QIN Bin-xin;QIU Chun(School of Automation, Nanjing Institute of Technology, Nanjing 211167, China)

机构地区：[1]南京工程学院自动化学院,江苏南京211167

出　　处：《南京工程学院学报（自然科学版）》2022年第1期37-42,共6页Journal of Nanjing Institute of Technology(Natural Science Edition)

摘　　要：为了更好地在复杂环境下跟踪到目标的运动轨迹,提出一种基于上下文感知的自适应目标跟踪算法.在满足实时性和精度的情况下,利用相关滤波联合全局上下文进行背景训练,利用目标及其上下文区域,实现协同跟踪.通过卡尔曼滤波估计出目标的运动方向,并在训练滤波器时将目标运动方向上的背景信息作为先验信息,降低非运动方向上背景样本的权重,增加运动方向上样本的权重.采用目标跟踪算法结合联合全局上下文算法训练尺度滤波器和位置滤波器,确定当前帧中目标的最优尺度.针对目标遮挡问题,以平均峰值相关能量作为遮挡判据指标对模型进行更新.在OTB—2015基准集上进行测试,该算法在多种复杂环境下也能较好地跟踪到目标运动轨迹,显著提高了相关滤波器的鲁棒性.A target tracking algorithm based on adaptive context awareness is proposed to track movements of a target in a complex environment.Background training is carried out by using correlation filtering and global context under the condition of real-time rate and accuracy.And cooperative tracking is completed via target and its context area.Kalman filter is used to estimate the target's motion direction with background information in the target's motion direction being prior information during filter training.The weight of the background samples in the non-motion direction is reduced,and the weight of the samples in the motion direction is increased.DSST(Discriminative Scale Space Tracking)algorithm is adopted to train the scale filter,and global context position filter to determine the optimal scale of the target in the current frame.Occlusion criterion APCE(Average Peak Correlation Energy)functions as a standard for updating the model.Through the tests on OTB—2015 benchmark set,the algorithm can also track the target trajectory in a variety of complex environments,which significantly improves the robustness of the correlation filter.

关 键 词：目标跟踪 视觉算法 遮挡判据 自适应 上下文感知 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]