学习时间变异抑制相关滤波的视频跟踪算法

A visual tracking algorithm for learning temporal andaberrance repressed correlation filters

目的:解决传统的基于判别相关滤波器的跟踪算法在遮挡、旋转、快速运动或低分辨率场景中容易出现检测异常或物体漂移等情况。方法:通过设计变异抑制正则项对检测阶段生成的响应图加以限制,使跟踪器能够有效抑制畸变。同时,构建时间正则项,利用前后帧的滤波器关系追踪目标,以有效缓解跟踪目标丢失问题。结果:采用精度图和成功率图作为评价标准,在标准数据集上与9种跟踪算法进行对比实验。实验结果表明,所提算法在遮挡、快速移动、背景干扰等检测异常情况下,对目标能够进行准确地跟踪。结论:通过设计变异抑制正则项和时间正则项,提高了传统的基于判别相关滤波跟踪器在复杂场景下的鲁棒性。

Aims:This papen aims to solve the problem of the abnormal detecting or object drifting of the traditional discriminant-correlation-filter based tracking algorithms in the case of occlusion,rotation,fast motion and low resolution.Methods:An aberrance repressed regularized term was designed to limit the responsing graphs generated in the detection phase.Meanwhile,a temporal regularized term was designed to prevent the loss of the tracked object by using the relationship between two continuous frames.Results:Compared with 9 tracking algorithms in the standard dataset,the proposed algorithm could track the target accurately under abnormal detections according to the precision and success rate of the graph.The proposed algorithm could accurately track the target in the case of abnormal detection.Conclusions:The design of the aberrance repressed regularized term and the temporal regularized term impoved the robustness of the traditional discriminant-correlation-filter in complex scenes.