摘要
针对目标在遮挡、尺度变化等复杂场景下易产生模型漂移问题,基于跟踪学习检测(TLD)框架提出一种结合基于网格的运动统计(GMS)检测和置信度判别的长时目标跟踪算法。首先在跟踪模块中采用快速判别尺度空间的相关滤波器(fDSST)作为跟踪器,利用位置滤波器和尺度滤波器对上一帧目标进行位置与尺度的判别,并依据TLD算法中跟踪模块与检测模块的独立性,将跟踪模块结果输入检测模块中,采用平均峰值相关能量(APCE)对模板更新进行置信度判别。在检测模块中先引入GMS网格运动统计作为检测器,使具有快速旋转不变性特征的ORB(OrientedFASTandRotatedBRIEF)算法对上一帧目标进行特征匹配,再利用网格运动统计对匹配结果进行过滤,实现目标位置的粗定位,依据预测位置对目标检测区域进行适当的动态缩减,最后使用级联分类器对目标进行精准定位。结果表明,本文提出的跟踪方法在有效防止模型漂移的情况下,大大提高了算法的跟踪速度,同时对目标遮挡、尺度变化及旋转等挑战环境也具有较好的准确性和鲁棒性。
In order to solve the problem of model drift in complex scenes such as occlusion and scale variation,this paper proposes a long-term target tracking algorithm based on TLD framework,which integrates GMS detection and confidence discrimination.First,in tracking module,the fast discriminating scale space correlation filter(fDSST)is used as the tracker,and the position filter and scale filter are used to distinguish the position and scale of the target in the previous frame.According to the independence of the tracking module and the detection module in the TLD algorithm,the results of the tracking module are input into the detection module,and the average peak-to-correlation energy(APCE)is used to determine the template update to judge the confidence.In the detection module,GMS grid motion statistics is used as the detector to make the ORB algorithm with fast rotation invariance feature to match the target in the previous frame,and then the grid motion statistics is used to filter the matching results to achieve the rough positioning of the target position,and the target detection area is reduced dynamically according to the prediction position.Finally,the cascaded classifier is used to locate the target accurately.The results show that the tracking method proposed in this paper can greatly improve the tracking speed of the algorithm while effectively preventing model drift,and has better accuracy and robustness to challenging environments such as target occlusion,scale variation and rotation.
作者
郭伟
杨琛
曲海成
邢宇哲
Guo Wei;Yang Chen;Qu Haicheng;Xing Yuzhe(School of Software,Liaoning Technical University,Huludao,Liaoning 125105,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2021年第4期155-166,共12页
Laser & Optoelectronics Progress
基金
国家自然科学基金(41701479)
辽宁省自然科学基金(20180550529)。
关键词
图像处理
目标跟踪
模型漂移
运动统计
动态缩减
image processing
target tracking
model drift
motion statistics
dynamic cutting
